Maven/internal/phraser/llmphraser.go

package phraser

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"log"
	"net/http"
	"os"
	"os/exec"
	"regexp"
	"strings"
	"sync"
	"syscall"
	"time"

	"github.com/kami/maven/internal/delivery"
	"github.com/kami/maven/internal/dialogue"
	"github.com/kami/maven/internal/loop"
	"github.com/kami/maven/internal/persona"
	"github.com/kami/maven/internal/router"
)

var listenRE = regexp.MustCompile(`listening on (https?://\S+)`)

// errEmptyResponse — the server answered and said nothing. Separate from a
// transport failure: the model is up and produced no tokens, which is still not
// an answer and must not score as one.
var errEmptyResponse = errors.New("phraser: empty response from the model")

type LLMPhraser struct {
	cfg    Config
	client *http.Client

	// tmpl — the hand-written Russian nudges. Default path for nudges; see
	// Config.LLMNudges. nil only if the template file failed to load.
	tmpl *NudgeTemplates

	// spawnCtx — the parent of every llama-server this phraser starts, i.e. the
	// daemon's own context. Deliberately NOT the per-request context of the call
	// that asked for a model swap: that one is cancelled the moment the request
	// returns, which would kill the model it had just loaded.
	spawnCtx context.Context
	cancel   context.CancelFunc

	// launch / probe — the two side effects of a swap, injectable so the swap
	// logic is testable without a real llama-server and a real model file.
	// launch is nil when this phraser does not own its server (NewLLMPhraserAt),
	// which is also what makes Swap refuse there.
	launch func(ctx context.Context, cfg Config) (backend, error)
	probe  func(ctx context.Context, base string) (string, error)

	// remote — the workstation model, when one is configured. Set once at wiring
	// time by UseRemote and read on every phrasing call. nil ⇒ every call goes to
	// the resident llama-server this phraser owns, which is the whole deploy
	// before a `workstation` block exists. See world.go.
	remote Remote

	// swapMu — single-flight around Swap. Held for the whole swap, including the
	// model load, so two concurrent swap requests can never both be loading.
	swapMu sync.Mutex

	// mu guards everything below: the live backend, the swap gate and the
	// in-flight request count. See acquire/quiesce in swap.go.
	mu        sync.Mutex
	be        backend
	live      liveModel
	swapping  bool
	inflight  int
	observers []func(baseURL string)
}

// liveModel — what is actually loaded right now. Distinct from Config, which
// stays immutable after construction: a swap changes these three fields and
// nothing else, so no reader of cfg (prompts, grammar, timeouts) races a swap.
type liveModel struct {
	ModelPath  string
	NGpuLayers int
	NCtx       int
}

type Config struct {
	ModelPath  string
	BinPath    string
	Listen     string
	NGpuLayers int
	NCtx       int
	Timeout    time.Duration

	// StartupTimeout bounds the wait for llama-server to print the address it
	// listens on. A config field and not a constant because the box may
	// legitimately need longer: a cold 1.7B loading off a spinning disk can
	// outrun a minute, and until this existed that returned "server did not
	// start within 60s" with no way to raise it.
	//
	// 0 ⇒ defaultStartupTimeout.
	StartupTimeout time.Duration

	// CacheRAMMiB bounds llama-server's prompt cache, which is what actually ate
	// this box. Measured on homesrv 2026-08-03: the server's own default limit is
	// 8192 MiB, it stores the full KV state of every idle slot it evicts (112 kiB
	// per token, so 166 MiB for one 1521-token prompt), and RSS climbed by that
	// much per distinct prompt until it hit 7.9 GB and half a gigabyte went to
	// swap. Weights are only 1.1 GB and mmapped, and -ngl 99 costs almost no RSS
	// because RADV keeps device memory outside the process.
	//
	// 0 ⇒ the flag is not passed and the server's own 8 GiB default applies. That
	// is the escape hatch for a llama-server too old to know --cache-ram, not a
	// recommendation. See docs/evals/2026-08-03-llama-prompt-cache.md.
	CacheRAMMiB int

	// ContextBlock renders the shared context block (who he is, how to
	// address him, the time) fresh for each turn. See internal/persona.
	// nil ⇒ no block, the prompts stand alone.
	ContextBlock func() string

	// LLMNudges puts the model back in charge of nudge wording.
	//
	// Off by default, and that is a deliberate deprecation of LLM-phrased
	// nudges: hand-written templates (nudges_ru_v1.json) word every nudge now.
	// A nudge has nothing to be creative about, and measured over many runs the
	// 0.8B broke the persona (formal "вы", plural imperatives, masculine
	// self-reference) and invented facts and units. Templates score 15/15 on the
	// nudge fixture, the model 11-13/15.
	//
	// The LLM path is kept, not deleted: flip this on to get it back. Chat,
	// query and reminder phrasing are untouched and still go through the model.
	LLMNudges bool

	// NoGrammar turns the GBNF constraint off (zero value ⇒ grammar ON).
	// The escape hatch exists because the target resident model — the
	// locally CPT'd Qwen3-1.7B — does not exist yet: if its chat template
	// ever fights the grammar, the fix should be a config flip on the
	// deploy box, not a code change and a rebuild.
	NoGrammar bool
}

func DefaultConfig(modelPath string) Config {
	return Config{
		ModelPath:  modelPath,
		BinPath:    "llama-server",
		Listen:     "127.0.0.1:0",
		NGpuLayers: -1,
		NCtx:       2048,
		// 512 MiB caps total RSS near 1 GB and still holds several recent prompts.
		CacheRAMMiB: 512,
		Timeout:     30 * time.Second,

		StartupTimeout: defaultStartupTimeout,
	}
}

func NewLLMPhraser(ctx context.Context, cfg Config) (*LLMPhraser, error) {
	ctx, cancel := context.WithCancel(ctx)
	p := &LLMPhraser{
		cfg:      cfg,
		client:   &http.Client{Timeout: cfg.Timeout},
		tmpl:     loadNudgeTemplates(),
		spawnCtx: ctx,
		cancel:   cancel,
		launch:   spawnLlamaServer,
		probe:    defaultProbe,
		live:     liveModel{ModelPath: cfg.ModelPath, NGpuLayers: cfg.NGpuLayers, NCtx: cfg.NCtx},
	}
	be, err := p.launch(ctx, cfg)
	if err != nil {
		cancel()
		return nil, err
	}
	p.be = be
	return p, nil
}

// NewLLMPhraserAt wires a phraser to a llama-server that someone else started
// and owns. It spawns nothing, so Close does not kill anything.
//
// This exists for the phrasing scorer (internal/phraser/eval), which must
// measure the phrasing against a shared llama-server without taking the model
// load hit per run or killing a server another process depends on. The daemon
// still uses NewLLMPhraser and still owns its own child process.
func NewLLMPhraserAt(baseURL string, cfg Config) *LLMPhraser {
	return &LLMPhraser{
		cfg:      cfg,
		client:   &http.Client{Timeout: cfg.Timeout},
		tmpl:     loadNudgeTemplates(),
		spawnCtx: context.Background(),
		cancel:   func() {},
		probe:    defaultProbe,
		// launch stays nil: we did not start this server, so we must not stop it.
		// Swap therefore refuses here (ErrSwapNotOwned) instead of killing a
		// server another process depends on.
		be:   borrowedBackend(strings.TrimSuffix(baseURL, "/")),
		live: liveModel{ModelPath: cfg.ModelPath, NGpuLayers: cfg.NGpuLayers, NCtx: cfg.NCtx},
	}
}

// loadNudgeTemplates loads the Russian nudge templates. A broken template file
// must not stop the daemon booting, so a failure logs and leaves the LLM path
// in charge of nudges.
func loadNudgeTemplates() *NudgeTemplates {
	nt, err := NewNudgeTemplates(nil)
	if err != nil {
		log.Printf("phraser: nudge templates unavailable, using the model: %v", err)
		return nil
	}
	return nt
}

// backend — one llama-server this phraser talks to. Two implementations: a
// llamaProc we spawned and must reap, and a borrowedBackend someone else owns.
type backend interface {
	BaseURL() string
	Close() error
}

// borrowedBackend — a server started and owned by someone else (the phrasing
// scorer's shared llama-server). Closing it is a no-op by construction.
type borrowedBackend string

func (b borrowedBackend) BaseURL() string { return string(b) }
func (b borrowedBackend) Close() error    { return nil }

// llamaProc — a llama-server child process plus the goroutine reading its
// stderr. Close kills and reaps it; see the Pdeathsig note in spawnLlamaServer.
type llamaProc struct {
	base   string
	cmd    *exec.Cmd
	cancel context.CancelFunc
	wg     sync.WaitGroup
}

func (l *llamaProc) BaseURL() string { return l.base }

func (l *llamaProc) Close() error {
	l.cancel()
	if l.cmd != nil && l.cmd.Process != nil {
		_ = l.cmd.Process.Kill()
		_ = l.cmd.Wait() // reap the process — without Wait, the child becomes a zombie
	}
	l.wg.Wait()
	return nil
}

// spawnLlamaServer starts one llama-server for cfg and waits until it says which
// address it is listening on. ctx owns the process lifetime, so it must be the
// daemon's context, not a request's.
func spawnLlamaServer(ctx context.Context, cfg Config) (backend, error) {
	ctx, cancel := context.WithCancel(ctx)
	p, err := startLlamaProc(ctx, cfg)
	if err != nil {
		cancel()
		return nil, err
	}
	p.cancel = cancel
	return p, nil
}

// llamaArgs is the command line for one resident server. It is a function and
// not an inline literal because kill-maven.sh's orphan sweep matches against
// this exact line, and a test pins the two together.
func llamaArgs(cfg Config) []string {
	args := []string{
		"-m", cfg.ModelPath,
		"--host", "127.0.0.1",
		"--port", extractPort(cfg.Listen),
		"-c", fmt.Sprintf("%d", cfg.NCtx),
		"-ngl", fmt.Sprintf("%d", cfg.NGpuLayers),
		"--no-webui",
	}
	if cfg.CacheRAMMiB > 0 {
		args = append(args, "--cache-ram", fmt.Sprintf("%d", cfg.CacheRAMMiB))
	}
	return args
}

// defaultStartupTimeout — the wait for llama-server's listen line when Config
// does not set one. A cold model load off disk is the slow part.
const defaultStartupTimeout = 60 * time.Second

func startLlamaProc(ctx context.Context, cfg Config) (*llamaProc, error) {
	startupTimeout := cfg.StartupTimeout
	if startupTimeout <= 0 {
		startupTimeout = defaultStartupTimeout
	}
	p := &llamaProc{}
	cmd := exec.CommandContext(ctx, cfg.BinPath, llamaArgs(cfg)...)
	// Pdeathsig: the kernel SIGKILLs llama-server the moment mavend dies — by
	// ANY means, including SIGKILL/OOM/panic where our Close() never runs. Without
	// it a hard-killed mavend orphans its llama-server (reparented to init, keeps
	// eating GPU/RAM); repeated dev restarts pile up orphans until the box OOMs.
	// Setpgid isolates it in its own process group so a stray Ctrl-C on the
	// terminal group doesn't half-kill it out from under us. (Linux-only, like
	// the rest of the daemon.)
	cmd.SysProcAttr = &syscall.SysProcAttr{Setpgid: true, Pdeathsig: syscall.SIGKILL}
	p.cmd = cmd

	// One pipe for both streams. llama.cpp writes its buffer sizes, KV-cache
	// layout and offload lines to stderr and its request log to stdout, and
	// stdout used to go nowhere at all — so nothing about the model's memory was
	// diagnosable from a running box. Both ends land in mavend's log now.
	pr, pw, err := os.Pipe()
	if err != nil {
		return nil, fmt.Errorf("llm: output pipe: %w", err)
	}
	cmd.Stdout = pw
	cmd.Stderr = pw

	if err := cmd.Start(); err != nil {
		pr.Close()
		pw.Close()
		return nil, fmt.Errorf("llm: start: %w", err)
	}
	// The child holds the only other reference to the write end. Dropping ours
	// is what makes the reader see EOF when the child dies.
	pw.Close()

	portCh := make(chan string, 1)
	errCh := make(chan error, 1)
	tail := &lineTail{}
	p.wg.Add(1)
	go func() {
		defer p.wg.Done()
		defer pr.Close()
		sc := bufio.NewScanner(pr)
		// llama.cpp prints one prompt per line and a prompt can be long.
		sc.Buffer(make([]byte, 0, 64*1024), 1024*1024)
		listening := false
		for sc.Scan() {
			line := sc.Bytes()
			log.Printf("llama: %s", line)
			if !listening {
				tail.add(string(line))
				if m := listenRE.FindSubmatch(line); len(m) > 1 {
					listening = true
					portCh <- string(m[1])
					close(portCh)
				}
			}
		}
		err := sc.Err()
		if err == nil {
			err = io.EOF
		}
		errCh <- err
	}()

	fail := func(err error) (*llamaProc, error) {
		_ = cmd.Process.Kill()
		_ = cmd.Wait()
		return nil, err
	}
	select {
	case addr := <-portCh:
		p.base = addr
		return p, nil
	case err := <-errCh:
		// The tail is the whole diagnosis when the server dies during load: bare
		// "EOF" never said which layer or which allocation it choked on.
		return fail(fmt.Errorf("llm: server output: %w; last output: %s", err, tail.String()))
	case <-ctx.Done():
		return fail(ctx.Err())
	case <-time.After(startupTimeout):
		return fail(fmt.Errorf("llm: server did not start within %s; last output: %s", startupTimeout, tail.String()))
	}
}

// lineTail keeps the last few startup lines so a server that dies before it
// listens can say why in the error, not just "EOF". Written by the reader
// goroutine and read by whoever gives up on startup, so it takes a lock.
type lineTail struct {
	mu    sync.Mutex
	lines []string
}

const lineTailMax = 12

func (t *lineTail) add(line string) {
	t.mu.Lock()
	defer t.mu.Unlock()
	t.lines = append(t.lines, line)
	if len(t.lines) > lineTailMax {
		t.lines = t.lines[len(t.lines)-lineTailMax:]
	}
}

func (t *lineTail) String() string {
	t.mu.Lock()
	defer t.mu.Unlock()
	if len(t.lines) == 0 {
		return "(no output)"
	}
	return strings.Join(t.lines, " | ")
}

// BaseURL is the llama-server this phraser talks to right now. It changes when
// the model is swapped, so callers that cache it must register an observer
// (OnSwap) rather than keeping the string forever.
func (p *LLMPhraser) BaseURL() string {
	p.mu.Lock()
	defer p.mu.Unlock()
	if p.be == nil {
		return ""
	}
	return p.be.BaseURL()
}

func (p *LLMPhraser) Close() error {
	p.cancel()
	p.mu.Lock()
	be := p.be
	p.be = nil
	p.mu.Unlock()
	if be != nil {
		return be.Close()
	}
	return nil
}

func (p *LLMPhraser) PhraseNudge(ctx context.Context, c loop.Candidate) (delivery.PhrasedNudge, error) {
	// Templates first — see Config.LLMNudges for why this is the default.
	if !p.cfg.LLMNudges && p.tmpl != nil {
		return p.tmpl.PhraseNudge(ctx, c)
	}
	prompt := buildNudgePrompt(c)
	resp, err := p.chat(ctx, prompt)
	if err != nil {
		return delivery.PhrasedNudge{}, err
	}
	body, mood, perr := parseResponseMood(resp)
	if perr != nil {
		// Truncated JSON. Not a nudge — use the plain Russian fallback.
		log.Printf("phraser: PhraseNudge: %v", perr)
		body, mood = "", ""
	}
	if body == "" {
		// fallback: try old body/summary format
		body, _ = parsePhrase(resp)
	}
	if body == "" {
		// The model said nothing usable. Say it in Russian anyway — this text
		// goes straight to a Russian piper voice, so the old "water — care"
		// fallback was unspeakable.
		body = fallbackNudge(c)
	}
	if mood == "" {
		mood = "neutral"
	}
	return delivery.PhrasedNudge{Candidate: c, Body: body, Summary: body, Mood: mood}, nil
}

// PhraseQuery prompts the LLM with the user's utterance and matching notes to
// compose a natural answer. On any LLM error it returns the fallback text —
// "вот что я нашла: <notes>", or "не знаю." with no notes — and the error
// together. The daemon uses the text and keeps the turn alive; a caller that is
// measuring counts the failure. Until Vikunja #397 the error was dropped, so a
// dead server scored as bad phrasing.
func (p *LLMPhraser) PhraseQuery(ctx context.Context, utterance string, notes []string) (string, error) {
	// Blank sources are no sources. A caller that hands over one empty string —
	// a page that fetched to nothing, a snippet trimmed away — used to take the
	// evidence branch and be told to answer from an empty list, which is the one
	// prompt guaranteed to make a small model fill the gap from memory.
	notes = nonEmpty(notes)
	if len(notes) == 0 {
		sys, prompt := p.knowledgePrompt(utterance)
		resp, err := p.chatWithSystem(ctx, sys, prompt, 768)
		if err != nil {
			return UnknownFallback(), fmt.Errorf("phrase query (knowledge): %w", err)
		}
		if resp == "" {
			return UnknownFallback(), errEmptyResponse
		}
		text, _, perr := parseResponseMood(resp)
		if perr != nil {
			return UnknownFallback(), fmt.Errorf("phrase query (knowledge): %w", perr)
		}
		if text != "" {
			return text, nil
		}
		return resp, nil
	}
	sys, prompt := p.evidencePrompt(utterance, notes)
	resp, err := p.chatWithSystem(ctx, sys, prompt, 768)
	text, _, perr := parseResponseMood(resp)
	if err != nil || perr != nil {
		// Read the notes out rather than ship a broken fragment.
		cause := err
		if cause == nil {
			cause = perr
		}
		return SourcesFallback(strings.Join(notes, "; ")),
			fmt.Errorf("phrase query (evidence): %w", cause)
	}
	if text != "" {
		return text, nil
	}
	return resp, nil
}

// PhraseChat uses the LLM to respond conversationally, building a multi-turn
// message array from dialogue history + the current user utterance. On any LLM
// error it returns both ChatFallback and the error, on the same rule as
// PhraseQuery: the fallback keeps the turn alive, the error stays visible.
func (p *LLMPhraser) PhraseChat(ctx context.Context, utterance string, history []dialogue.Turn) (string, error) {
	sys := chatSystemPrompt(p.cfg.ContextBlock)
	msgs := []chatMsg{
		{Role: "system", Content: sys},
	}
	// Combine history and current utterance into one user message.
	// Some model chat templates (Ministral, etc.) reject consecutive user turns.
	var combined string
	for _, t := range history {
		combined += t.Text + "\n"
	}
	combined += utterance
	msgs = append(msgs, chatMsg{Role: "user", Content: strings.TrimSpace(combined)})

	resp, err := p.chatWithMessages(ctx, msgs, 768)
	if err != nil {
		return ChatFallback(), fmt.Errorf("phrase chat: %w", err)
	}
	text, _, perr := parseResponseMood(resp)
	if perr != nil {
		return ChatFallback(), fmt.Errorf("phrase chat: %w", perr)
	}
	if text != "" {
		return text, nil
	}
	// fallback: plain text without JSON
	if i := strings.IndexByte(resp, '\n'); i >= 0 {
		resp = resp[:i]
	}
	return strings.TrimSpace(resp), nil
}

// chatSystemPrompt returns the system prompt for conversational chat.
// Prepends the shared context block when the phraser has one.
func chatSystemPrompt(block func() string) string {
	// No self-introduction here: the persona block prepended one line above
	// already says who she is, same as router.KnowledgePrompt.
	//
	// The grammar examples used to be full clauses: ("я подумала", "я рада")
	// for her, ("ты сказал", "ты забыл") for him. A 1.7B copies those rather
	// than generalising from them. Observed on the box 2026-08-01: all three
	// chat replies in one session opened with "Я подумала, что ...", and one
	// ended "...немного тревожусь. ты сказал" — the second example pasted onto
	// the end of a finished sentence, which reads as a truncation but is not.
	//
	// So: contrastive pairs instead of usable openers. "рада, не рад" states
	// the rule as a correction, and short predicatives do not hand the model a
	// sentence frame to start with. The him-examples are gone entirely; the
	// "ты" instruction carries that on its own and those two produced the
	// worst output. The last line says outright not to echo the instructions,
	// because a small model will otherwise treat any quoted string as licence.
	//
	// Amended the same day: with the openers gone the tic went with them, but
	// "не забыл ли я" appeared — masculine, about herself. The old "я подумала"
	// had been suppressing that by accident, being a feminine past tense the
	// model could copy. Two short predicatives are not enough signal on their
	// own, so the rule is now stated as morphology (-ла) rather than as a pair
	// of words. A suffix rule generalises where an example only gets copied.
	base := `Ты разговариваешь с хозяином.

О себе — в женском роде: "рада", не "рад"; "поняла", не "понял". Все свои глаголы в прошедшем времени оканчивай на -ла: сделала, забыла, записала, подумала. Он мужчина: обращайся к нему на "ты", в мужском роде. Никогда не "вы"/"ваш" и никогда "он"/"его" — ты говоришь ему, а не о нём.

Отвечай по-русски, коротко: одна-три фразы, живым языком. Ты доброжелательная, тебе интересно, но чувства не изображай. Не повторяй формулировки из этой инструкции — отвечай своими словами.

Отвечай ТОЛЬКО одним объектом JSON: {"response": "...", "mood": "neutral"}. В "response" — твой ответ. В "mood" — ровно одно из: neutral, happy, thinking, tired, confused.`
	return persona.Prepend(block, base)
}

// chatWithMessages sends a full message array (system + history + current) to
// the LLM completion endpoint. Like chatWithSystem but for an arbitrary message
// slice — the caller owns the system prompt placement.
func (p *LLMPhraser) chatWithMessages(ctx context.Context, msgs []chatMsg, maxTokens int) (string, error) {
	// Same silent preference as chatWithSystem, when the array is the shape
	// llm.Req can carry: one system turn and one user turn. PhraseChat already
	// folds the history into a single user message (some chat templates reject
	// consecutive user turns), so today that is every call. A longer array goes
	// to the resident model rather than get flattened here, because flattening a
	// conversation is a decision its owner should make.
	if len(msgs) == 2 && msgs[0].Role == "system" && msgs[1].Role == "user" {
		if out, ok := p.remoteChat(ctx, msgs[0].Content, msgs[1].Content, maxTokens); ok {
			return out, nil
		}
	}
	base, release, err := p.acquire()
	if err != nil {
		return "", err
	}
	defer release()
	req := chatReq{
		Messages:    msgs,
		Temperature: 0.7,
		MaxTokens:   maxTokens,
		Grammar:     p.grammar(),
	}
	body, err := json.Marshal(req)
	if err != nil {
		return "", fmt.Errorf("llm: marshal: %w", err)
	}
	httpReq, err := http.NewRequestWithContext(ctx, "POST", base+"/v1/chat/completions", bytes.NewReader(body))
	if err != nil {
		return "", fmt.Errorf("llm: request: %w", err)
	}
	httpReq.Header.Set("Content-Type", "application/json")
	resp, err := p.client.Do(httpReq)
	if err != nil {
		return "", fmt.Errorf("llm: post: %w", err)
	}
	defer resp.Body.Close()
	raw, err := io.ReadAll(resp.Body)
	if err != nil {
		return "", fmt.Errorf("llm: read: %w", err)
	}
	if resp.StatusCode != 200 {
		return "", fmt.Errorf("llm: status %d: %s", resp.StatusCode, strings.TrimSpace(string(raw)))
	}
	var cr chatResp
	if err := json.Unmarshal(raw, &cr); err != nil {
		return "", fmt.Errorf("llm: parse: %w", err)
	}
	if len(cr.Choices) == 0 {
		return "", fmt.Errorf("llm: no choices in response")
	}
	content := cr.Choices[0].Message.Content
	if content == "" {
		content = cr.Choices[0].Message.ReasoningContent
	}
	log.Printf("llm raw content: %q", content)
	return stripThink(content), nil
}

func (p *LLMPhraser) PhraseReminder(ctx context.Context, d loop.ReminderDecision) (delivery.PhrasedReminder, error) {
	text := extractReminderText(d.Reminder.Payload)
	if text == "" {
		text = "reminder"
	}

	// Russian, like the other two prompts (Vikunja #404). Asking a model for a
	// Russian reply in English is asking it to switch languages mid-prompt,
	// and a 1.7B sometimes answers in the language it was asked in. The
	// persona rules and the JSON contract are not repeated here: this call
	// goes through chat(), so nudgeSystem already states both, and a second
	// statement of the same contract is one more thing that can drift.
	prompt := fmt.Sprintf(
		`Он поставил напоминание: "%s". Скажи это своими словами, коротко и мягко — одно предложение.`,
		text,
	)
	resp, err := p.chat(ctx, prompt)
	if err != nil {
		return delivery.PhrasedReminder{}, err
	}
	body, mood, perr := parseResponseMood(resp)
	if perr != nil {
		// Truncated JSON. Fall through to the reminder's own text.
		log.Printf("phraser: PhraseReminder: %v", perr)
		body, mood = "", ""
	}
	if body == "" {
		// fallback: try old body/summary format
		body, _ = parsePhrase(resp)
	}
	if body == "" {
		body = text
	}
	if mood == "" {
		mood = "neutral"
	}
	summary := body
	if len(summary) > 60 {
		summary = summary[:57] + "..."
	}
	return delivery.PhrasedReminder{Decision: d, Body: body, Summary: summary, Mood: mood}, nil
}

type chatMsg struct {
	Role    string `json:"role"`
	Content string `json:"content"`
}

type chatReq struct {
	Model       string    `json:"model"`
	Messages    []chatMsg `json:"messages"`
	Temperature float64   `json:"temperature"`
	MaxTokens   int       `json:"max_tokens"`
	// Grammar is llama-server's `grammar` field (GBNF). Same wiring as
	// internal/llm.Req.Grammar. Empty ⇒ unconstrained sampling.
	Grammar string `json:"grammar,omitempty"`
}

// responseGrammar — GBNF constraining the model to the documented phrasing
// contract and nothing else: {"response": "<text>", "mood": "<enum>"}.
//
// Without it a 0.8B answers roughly one chat turn in three with open reasoning
// as plain text ("Thinking Process:" …), which no tag-stripper can remove and
// which eats the token budget before the JSON closes. Modelled on
// routeGrammar in internal/router/llmrouter.go so the two read alike.
//
// text accepts ANY codepoint except the two JSON must escape — the replies are
// Russian, so an ASCII-only rule would make every reply empty. The escape rule
// is what lets the model close a string it opened with a quote inside. Length
// is bounded so a repetition loop truncates the field, not the JSON object.
//
// That bound was 400 and 400 was too tight. Measured against Qwen3.5-0.8B: on
// "почему гром слышно позже молнии?" the reply came back exactly 400 characters
// long, cut mid-word ("Нужно записать и,"), at every token cap from 256 to 2048.
// So the token cap was never what stopped it — this rule was. 1000 characters is
// roughly six Russian sentences, still short enough to stop a repetition loop.
const responseGrammar = `
root   ::= "{" ws "\"response\"" ws ":" ws string ws "," ws "\"mood\"" ws ":" ws mood ws "}"
mood   ::= "\"neutral\"" | "\"happy\"" | "\"thinking\"" | "\"tired\"" | "\"confused\""
string ::= "\"" ([^"\\] | "\\" ["\\/bfnrt]){0,1000} "\""
ws     ::= [ \t\n]*
`

// ResponseGrammar exposes responseGrammar to the other callers that emit the
// same {"response","mood"} contract — cmd/mavend's reactive replier, which is
// parsed by the same two fields. One definition, so the two cannot drift.
const ResponseGrammar = responseGrammar

// grammar returns the GBNF to attach to a phrasing request, or "" when the
// operator turned it off.
func (p *LLMPhraser) grammar() string {
	if p.cfg.NoGrammar {
		return ""
	}
	return responseGrammar
}

type chatResp struct {
	Choices []struct {
		Message struct {
			Content          string `json:"content"`
			Reasoning        string `json:"reasoning"`
			ReasoningContent string `json:"reasoning_content"`
		} `json:"message"`
	} `json:"choices"`
}

func (p *LLMPhraser) chat(ctx context.Context, userPrompt string) (string, error) {
	return p.chatWithSystem(ctx, p.systemPrompt(), userPrompt, 512)
}

func (p *LLMPhraser) chatWithSystem(ctx context.Context, system, user string, maxTokens int) (string, error) {
	// The workstation model first when it will take work, and silently: every
	// caller of this helper is on the silent half of the degradation rule. It
	// answering is not news, and it being asleep is not news either.
	if out, ok := p.remoteChat(ctx, system, user, maxTokens); ok {
		return out, nil
	}
	base, release, err := p.acquire()
	if err != nil {
		return "", err
	}
	defer release()
	req := chatReq{
		Messages: []chatMsg{
			{Role: "system", Content: system},
			{Role: "user", Content: user},
		},
		Temperature: 0.7,
		MaxTokens:   maxTokens,
		Grammar:     p.grammar(),
	}
	body, err := json.Marshal(req)
	if err != nil {
		return "", fmt.Errorf("llm: marshal: %w", err)
	}

	httpReq, err := http.NewRequestWithContext(ctx, "POST", base+"/v1/chat/completions", bytes.NewReader(body))
	if err != nil {
		return "", fmt.Errorf("llm: request: %w", err)
	}
	httpReq.Header.Set("Content-Type", "application/json")

	resp, err := p.client.Do(httpReq)
	if err != nil {
		return "", fmt.Errorf("llm: post: %w", err)
	}
	defer resp.Body.Close()

	raw, err := io.ReadAll(resp.Body)
	if err != nil {
		return "", fmt.Errorf("llm: read: %w", err)
	}
	if resp.StatusCode != 200 {
		return "", fmt.Errorf("llm: status %d: %s", resp.StatusCode, strings.TrimSpace(string(raw)))
	}

	var cr chatResp
	if err := json.Unmarshal(raw, &cr); err != nil {
		return "", fmt.Errorf("llm: parse: %w", err)
	}
	if len(cr.Choices) == 0 {
		return "", fmt.Errorf("llm: no choices in response")
	}
	content := cr.Choices[0].Message.Content
	if content == "" {
		content = cr.Choices[0].Message.ReasoningContent
	}
	return stripThink(content), nil
}

// nudgeSystem — the phrasing contract for nudges.
//
// Written as filled-in examples, not as a schema with "..." in it. A 0.8B
// copies whatever sits in the response slot, so a literal placeholder there
// teaches it to answer with the placeholder. Measured: 7/15 nudges came back
// as "..." before this. See docs/evals/2026-07-31-phrasing.md.
//
// Russian only, feminine self-reference, second person masculine (the owner is
// a man). She talks TO him, informally, singular — never "вы", never "он".
// One short sentence — the nudge is spoken aloud.
//
// What the ban on обращения forbids is pet names ("дорогой", "милый"), not his
// name: "Ками, ноутбук на трёх процентах" is exactly how she talks, and the
// unqualified word read as forbidding that too. Hence "ласковые обращения".
//
// The examples also never claim a physical act. She has no hands and no smart
// plug — she can tell him the battery is at three percent, she cannot put the
// laptop on charge. An example that says she did teaches the model to invent
// actions Maven never took, which is worse than a missing nudge.
const nudgeSystem = `Ты — Maven, домашняя ассистентка. О себе говоришь в женском роде ("я проверила", "я записала"). Владелец — мужчина, обращайся к нему в мужском роде ("ты пил", "ты забыл").
Говоришь с ним на "ты", в единственном числе ("выпей", "встань"). Никогда не "вы"/"вас"/"ваш" и никогда "он"/"его" — ты говоришь ему, а не о нём.

Пиши ОДНО короткое напоминание по-русски: не больше 120 символов и не больше 16 слов. Только по делу.

Запрещено: ласковые обращения ("дорогой", "милый"), эмодзи, извинения ("прости", "извини"), вопросы о самочувствии, похвала, больше одного восклицательного знака, английские слова кроме имён сервисов.

Отвечай ТОЛЬКО одним объектом JSON с полями "response" и "mood".
"response" — сам текст напоминания.
"mood" — ровно одно из: neutral, happy, thinking, tired, confused.

Так выглядит правильный ответ по форме. Темы здесь посторонние — их в запросе не будет:
{"response": "Стиральная машина закончила. Развесь бельё.", "mood": "neutral"}
{"response": "Ками, ноутбук на трёх процентах. Поставь его на зарядку.", "mood": "confused"}

Это примеры ФОРМЫ, а не темы. Пиши только про ту ситуацию, которую тебе дали в запросе. Не копируй примеры и никогда не пиши "..." в поле response.`

func (p *LLMPhraser) systemPrompt() string {
	return persona.Prepend(p.cfg.ContextBlock, nudgeSystem)
}

// knowledgePrompt — the no-sources branch: a world question, answered from
// weights alone. The system prompt is the single tested source in
// router.KnowledgePrompt.
//
// Split out of PhraseQuery so PhraseWorld sends the workstation model the same
// bytes the resident model gets. Prompt parity across two models is a stated
// constraint (CLAUDE.md), and two copies of a prompt is how it stops holding.
func (p *LLMPhraser) knowledgePrompt(utterance string) (sys, user string) {
	return persona.Prepend(p.cfg.ContextBlock, router.KnowledgePrompt()),
		fmt.Sprintf("Пользователь спрашивает: \"%s\".", utterance)
}

// evidencePrompt — the sources branch: read these, add nothing. Shared with
// PhraseWorld for the same reason as knowledgePrompt.
func (p *LLMPhraser) evidencePrompt(utterance string, notes []string) (sys, user string) {
	return p.querySystemPrompt(), fmt.Sprintf(
		"Он спрашивает: \"%s\"\n\nИсточники:\n%s\nОтветь ему коротко и своими словами, опираясь только на эти источники. Если ответа в них нет — так и скажи.",
		utterance, evidenceBlock(notes),
	)
}

// querySystemPrompt returns the system prompt for the evidence branch of
// PhraseQuery. Prepends the configured persona when set.
//
// Evidence-first, and that is the whole point of this prompt. Every source that
// reaches PhraseQuery with something in hand — his notes, a stored fact, a page,
// a live search, a ZIM article — arrives as numbered sources, and the model's
// job here is to READ them, not to recall. A 1.7B asked a world question
// answers from its weights with total confidence and no signal that it is
// guessing; that is how "Война и мир" got Левитан as its author. The rule that
// prevents it is stated three ways, because one way did not hold: answer from
// the sources, say plainly when they do not answer, add nothing of your own.
//
// It no longer says "заметки". The sources are not always his notes, and
// calling a Wikipedia paragraph his note both misleads him and licenses the
// model to blur where an answer came from.
//
// No self-introduction here: the persona block prepended one line above already
// says who she is, same as router.KnowledgePrompt.
//
// The opener is deliberate and stays: the fixed prefix is what marks the answer
// as a lookup rather than as something she knows. The grammar examples are not
// deliberate — same defect chatSystemPrompt had, where a 1.7B copies a quoted
// word instead of generalising from it. Stated as morphology instead.
func (p *LLMPhraser) querySystemPrompt() string {
	base := "Ты отвечаешь ему по источникам, которые тебе дали. Отвечай ТОЛЬКО по ним: всё, что ты говоришь, должно быть написано в источниках. " +
		"Если ответа в них нет — так и скажи и на этом остановись; не добавляй ничего из своих знаний и не догадывайся. " +
		"Не приплетай прошлые реплики разговора. " +
		"Отвечай по-русски, коротко и своими словами, начинай с \"вот что я нашла: \". О себе — в женском роде, глаголы в прошедшем времени с окончанием -ла. Он мужчина, обращайся к нему на \"ты\". Отвечай ТОЛЬКО одним объектом JSON: {\"response\": \"...\", \"mood\": \"neutral\"}."
	return persona.Prepend(p.cfg.ContextBlock, base)
}

// evidenceBlock renders the sources for the evidence branch of PhraseQuery.
//
// Numbered lines, one source each, rather than the quoted semicolon-joined
// string this used to build. Two reasons, both measured on small models: a
// numbered list survives being long, where a run-on quoted string blurs into
// one claim the model then merges; and the numbering gives it something to
// answer FROM, which is what makes "этого в источниках нет" reachable at all.
func evidenceBlock(sources []string) string {
	var b strings.Builder
	for i, s := range sources {
		fmt.Fprintf(&b, "[%d] %s\n", i+1, s)
	}
	return b.String()
}

// nonEmpty drops blank sources and trims the rest, without touching the
// caller's slice.
func nonEmpty(sources []string) []string {
	out := make([]string, 0, len(sources))
	for _, s := range sources {
		if s = strings.TrimSpace(s); s != "" {
			out = append(out, s)
		}
	}
	return out
}

// ruleTopics — Russian gloss for each built-in rule name. The rule names are
// English identifiers; a 0.8B asked to nudge about "netdata_critical" writes
// about nothing. The daemon knows what its own rules mean, so it says so.
var ruleTopics = map[string]string{
	"water":            "он давно не пил воду",
	"meal":             "он давно не ел",
	"break":            "он давно без перерыва, пора встать и размяться",
	"service_down":     "сервис не отвечает, лежит",
	"netdata_critical": "критический алярм в netdata, проблема с диском или местом",
}

// ruleKeywords — the word the message must contain. The 0.8B drifts to
// whatever topic it saw last unless the required word is named outright.
var ruleKeywords = map[string]string{
	"water":            "воду",
	"meal":             "поешь",
	"break":            "перерыв",
	"service_down":     "сервис",
	"netdata_critical": "диск",
}

// ruleTopic turns a rule name into a Russian description of the situation.
// "routine:зарядка" and "morning:утро" carry their own Russian suffix.
func ruleTopic(rule string) string {
	if t, ok := ruleTopics[rule]; ok {
		return t
	}
	if i := strings.IndexByte(rule, ':'); i > 0 && i+1 < len(rule) {
		switch rule[:i] {
		case "morning":
			return "утро, пора начать день: " + rule[i+1:]
		default:
			return "пора сделать по распорядку: " + rule[i+1:]
		}
	}
	return rule
}

// ruleKeyword — the word the nudge must contain, or "" when the rule name's
// own Russian suffix already is that word.
func ruleKeyword(rule string) string {
	if k, ok := ruleKeywords[rule]; ok {
		return k
	}
	if i := strings.IndexByte(rule, ':'); i > 0 && i+1 < len(rule) {
		return rule[i+1:]
	}
	return ""
}

// ruDur — duration in Russian. humanDur is English and its output was landing
// verbatim in the message.
func ruDur(d time.Duration) string {
	if d < 0 {
		d = 0
	}
	h, m := int(d.Hours()), int(d.Minutes())%60
	switch {
	case h >= 2:
		return fmt.Sprintf("%d ч", h)
	case h == 1 && m >= 30:
		return "полтора часа"
	case h == 1:
		return "час"
	default:
		return fmt.Sprintf("%d мин", m)
	}
}

// fallbackNudge — plain Russian for when the model returns nothing parseable.
var fallbackNudges = map[string]string{
	"water":            "Ты давно не пил воду.",
	"meal":             "Ты давно не ел, поешь.",
	"break":            "Пора сделать перерыв.",
	"service_down":     "Сервис не отвечает.",
	"netdata_critical": "Критический алярм: проверь диск.",
}

func fallbackNudge(c loop.Candidate) string {
	if down := loop.DownServices(c.State); len(down) > 0 {
		return "Не отвечает: " + strings.Join(down, ", ") + "."
	}
	if s, ok := fallbackNudges[c.Rule.Name]; ok {
		return s
	}
	if kw := ruleKeyword(c.Rule.Name); kw != "" {
		return "Напоминаю: " + kw + "."
	}
	return "Напоминаю о деле."
}

func buildNudgePrompt(c loop.Candidate) string {
	var ctxParts []string
	ctxParts = append(ctxParts, "Ситуация: "+ruleTopic(c.Rule.Name))
	if f, ok := c.State.Facts[c.Rule.Name]; ok && f.Key != "" && f.Key != c.Rule.Name {
		ctxParts = append(ctxParts, "Что именно: "+f.Key)
	}
	if down := loop.DownServices(c.State); len(down) > 0 {
		// The names come from the same helper the rule fired on, so the model
		// is never handed a service that is actually up.
		ctxParts = append(ctxParts, "Какие сервисы лежат: "+strings.Join(down, ", "))
	}
	if d, ok := c.State.Since(c.Rule.Name); ok {
		ctxParts = append(ctxParts, "Прошло: "+ruDur(d))
	}
	switch sevLabel(c.Severity) {
	case "alarm":
		ctxParts = append(ctxParts, "Срочно, скажи прямо.")
	case "ops":
		ctxParts = append(ctxParts, "Это про сервер, не про здоровье.")
	}
	tail := "Напиши напоминание про эту ситуацию. Одно предложение, по-русски, в JSON."
	if kw := ruleKeyword(c.Rule.Name); kw != "" {
		// Last line on purpose: a 0.8B weights the end of the prompt hardest,
		// and without the required word it drifts back to the examples.
		tail += " Ответ ДОЛЖЕН содержать слово «" + kw + "»."
	}

	return strings.Join(ctxParts, "\n") + "\n\n" + tail
}

type responseMood struct {
	Response string `json:"response"`
	Mood     string `json:"mood"`
}

// errBrokenJSON — the model started a JSON object and never finished it.
// That is a failed generation, not a reply. Callers must use their fallback.
var errBrokenJSON = fmt.Errorf("phraser: model output starts as JSON but does not parse")

// parseResponseMood extracts {"response","mood"} from LLM output, tolerant
// of thinking tokens and extra text before/after the JSON block.
//
// Three outcomes:
//   - parsed fine → the fields, nil error.
//   - output never looked like JSON → ("", "", nil). The caller may ship it
//     as-is; small models sometimes answer in bare prose and that is fine.
//   - output starts with "{" but does not parse → errBrokenJSON. The grammar
//     guarantees a valid *prefix*, so a generation that hits the token cap
//     mid-object comes back as a fragment like `{` or `{\n  "`. Shipping that
//     as a reply is the bug this error exists to stop.
func parseResponseMood(raw string) (response, mood string, err error) {
	cleaned := strings.TrimSpace(raw)
	start := strings.Index(cleaned, "{")
	end := strings.LastIndex(cleaned, "}")
	if start < 0 || end < 0 || end <= start {
		if strings.HasPrefix(cleaned, "{") {
			return "", "", errBrokenJSON
		}
		return "", "", nil
	}
	var parsed responseMood
	if e := json.Unmarshal([]byte(cleaned[start:end+1]), &parsed); e != nil {
		if strings.HasPrefix(cleaned, "{") {
			return "", "", errBrokenJSON
		}
		return "", "", nil
	}
	return parsed.Response, parsed.Mood, nil
}

func parsePhrase(raw string) (body, summary string) {
	cleaned := strings.TrimSpace(raw)
	start := strings.Index(cleaned, "{")
	end := strings.LastIndex(cleaned, "}")
	if start < 0 || end < 0 || end <= start {
		return "", ""
	}
	var parsed struct {
		Body    string `json:"body"`
		Summary string `json:"summary"`
	}
	if err := json.Unmarshal([]byte(cleaned[start:end+1]), &parsed); err != nil {
		return "", ""
	}
	return parsed.Body, parsed.Summary
}

func extractPort(listen string) string {
	_, port, _ := strings.Cut(listen, ":")
	if port == "" {
		return "0"
	}
	return port
}

// stripThink removes the <think> block that Thinking-variant models emit
// before the actual response. No-op when no think block is present.
func stripThink(s string) string {
	if i := strings.LastIndex(s, "</think>"); i >= 0 {
		s = strings.TrimSpace(s[i+8:])
	}
	return s
}