Maven/internal/router/router.go

package router

import (
	"context"
	"log"
	"time"
)

// Config — wires the cascade. Build via New; a zero-value Router is unusable.
type Config struct {
	// Grammars — stage-0 exact-match rules. DefaultGrammars(actMatcher) wires
	// the wake-word act fast path; the daemon may append more.
	Grammars []Grammar
	// Classifier — stage-1 nearest-centroid classifier. Must be seeded with
	// ~10 examples/intent at bootstrap (per spec) before free-form routing
	// is trustworthy; until then Route returns ErrNoIntents on free-form input.
	Classifier *Classifier
	// Extractor — stage-2 per-intent slot extraction. Any nil sub-parser just
	// leaves the corresponding Has* flag false for that intent.
	Extractor Extractor
	// Threshold — stage-3 confidence gate. Below ⇒ Clarify, don't guess. The
	// spec leaves this open (defines how often maven asks vs guesses on free-
	// form input; the whole reactive mvp feel rides on it). The daemon sets it.
	Threshold float64
	// LLM — optional agentic router. When set, Route consults it after stage-0
	// and before the classifier cascade, classifying the utterance via a
	// grammar-constrained call to the resident model (Qwen3-1.7B). On any
	// error/parse failure, falls through to the classifier (never fails the
	// turn on the model).
	LLM *LLMRouter
}

// Router — the deterministic cascade. Route never guesses: stage 0 wins
// outright, stage 1 scores, stage 2 extracts, stage 3 gates. The SLM only
// phrases the reply — it never owns the route.
type Router struct {
	grammars   []Grammar
	classifier *Classifier
	extractor  Extractor
	threshold  float64
	llm        *LLMRouter
}

func New(cfg Config) *Router {
	return &Router{
		grammars:   cfg.Grammars,
		classifier: cfg.Classifier,
		extractor:  cfg.Extractor,
		threshold:  cfg.Threshold,
		llm:        cfg.LLM,
	}
}

// Route — the cascade: stage 0 (exact match) → 1 (classify) → 2 (extract) →
// 3 (confidence gate).
//
//	Stage 0 wins outright: returns at confidence 1.0, no classifier.
//	Otherwise the classifier scores every intent; the best wins; slots are
//	extracted for that intent. If the winning score < threshold the Decision
//	is flagged Clarify (the daemon asks rather than guesses — same shape as
//	since(key)==null → don't fire: a misrouted fact is a confident wrong write,
//	worse than a gap).
func (r *Router) Route(ctx context.Context, utterance string, now time.Time) (Decision, error) {
	// stage 0 — exact match / grammar. First match wins; grammars are ordered.
	// Grammars like time/date/reminder don't expect a wake-word prefix, but
	// the STT often includes one (transcribed phonetically, any script) — try
	// the wake-stripped utterance too so those grammars still fire.
	stripped, hadWake := StripWakeToken(utterance)
	for _, g := range r.grammars {
		m := g.Pattern.FindStringSubmatch(utterance)
		if m == nil && hadWake {
			m = g.Pattern.FindStringSubmatch(stripped)
		}
		if m == nil {
			continue
		}
		d, ok := g.Build(m)
		if !ok {
			continue // grammar matched shape but not content → fall through
		}
		d.Utterance = utterance
		return d, nil
	}

	// 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
			r.fillSlots(ctx, &d, now)
			r.gateLLMDecision(&d)
			return d, nil
		} else if err != nil {
			log.Printf("router: llm route fell back to classifier: %v", err)
		}
	}

	// stage 1 — intent classifier.
	results, err := r.classifier.Classify(ctx, utterance)
	if err != nil {
		return Decision{}, err
	}
	best := results[0]

	// stage 2 — slot extraction for the winning intent.
	d := Decision{
		Utterance:  utterance,
		Stage:      2,
		Intent:     best.Intent,
		Confidence: best.Score,
		Slots:      r.extractor.Extract(ctx, best.Intent, utterance, now),
	}

	// stage 3 — confidence gate. Below threshold ⇒ clarify, don't guess.
	if d.Confidence < r.threshold {
		d.Stage = 3
		d.Clarify = true
	}
	return d, nil
}

// fillSlots — run stage-2 extraction on an LLM decision and fill only the slots
// the model left empty. The LLM wins where it answered: it saw the sentence, the
// parsers are keyword tables. Extraction covers what the model cannot produce at
// all — a parsed reminder time and an allowlist fn.
//
// If a reminder still has no time, leave it missing. The daemon then says it
// could not read the time; inventing one would set a wrong alarm.
func (r *Router) fillSlots(ctx context.Context, d *Decision, now time.Time) {
	ex := r.extractor.Extract(ctx, d.Intent, d.Utterance, now)
	if !d.Slots.HasTime && ex.HasTime {
		d.Slots.Time, d.Slots.HasTime = ex.Time, ex.HasTime
	}
	if !d.Slots.HasKey && ex.HasKey {
		d.Slots.Key, d.Slots.Value, d.Slots.HasKey = ex.Key, ex.Value, ex.HasKey
	}
	if !d.Slots.HasFn && ex.HasFn {
		d.Slots.Fn, d.Slots.Args, d.Slots.HasFn = ex.Fn, ex.Args, ex.HasFn
	}
	// For an act the model returns the verb in Text ("restart nginx"), which is
	// often cleaner than the raw utterance ("maven, could you restart nginx").
	// Try it too when the utterance did not match the allowlist.
	if d.Intent == IntentAct && !d.Slots.HasFn && r.extractor.Acts != nil &&
		d.Slots.Text != "" && d.Slots.Text != d.Utterance {
		if fn, args, ok := r.extractor.Acts.Match(d.Slots.Text); ok {
			d.Slots.Fn, d.Slots.Args, d.Slots.HasFn = fn, args, true
		}
	}
	if d.Slots.Text == "" {
		d.Slots.Text = ex.Text
	}
	// Stage stays 1: it says who decided the route, and that was the LLM.
}

// gateLLMDecision — stage 3 for the LLM path (Vikunja #359). This used to be
// the classifier's job alone (see the threshold check at the bottom of
// Route): the LLM branch returned straight from fillSlots and never touched
// r.threshold at all, so a hardcoded Confidence: 1.0 in llmrouter.go could
// never gate. Two more structural holes are checked here, after fillSlots
// has had a chance to fill them from the deterministic parsers — checking
// before fillSlots would flag e.g. every keyless fact the fact parser goes
// on to resolve (TestLLMFactGetsKeyFromParser):
//   - a fact with no key even after the parser tried — nothing to write, or
//     worse, a confident write under the wrong key;
//   - an act that never resolved to an allowlisted fn — a confident guess
//     here means either silently doing nothing or, if the daemon is lax,
//     running something never on the allowlist. Don't guess; ask.
//
// Anything below threshold gets the exact same Clarify=true treatment the
// classifier path already produces — same field, same daemon-side consumer
// (cmd/mavend/clarify.go), nothing new to wire.
func (r *Router) gateLLMDecision(d *Decision) {
	if d.Intent == IntentFact && !d.Slots.HasKey && d.Confidence > llmThinConfidence {
		d.Confidence = llmThinConfidence
	}
	if d.Intent == IntentAct && !d.Slots.HasFn && d.Confidence > llmThinConfidence {
		d.Confidence = llmThinConfidence
	}
	if d.Confidence < r.threshold {
		d.Clarify = true
	}
}

// CorrectMisroute — the user corrected a bad classification. Appends a new
// example for the corrected intent (append-only — grows the classifier, no
// retrain). Same shape as nudges.outcome tuning cooldowns: more reliable over
// time, introspectable, no model surgery.
func (r *Router) CorrectMisroute(ctx context.Context, utterance string, corrected Intent) error {
	return r.classifier.AddExample(ctx, corrected, utterance)
}