Fix kill-maven.sh so it actually kills llama-server

The MODEL default was LFM2, but the deploy runs Qwen3.5-0.8B, so the
pkill pattern matched nothing and the server survived every kill.
Now matches any llama-server serving a .gguf, so changing the model in
deploy/mavend.json cannot break the script again. MODEL still narrows it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ

cmd/mavend/llmrouter_flag_test.go

@@ -0,0 +1,28 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"github.com/kami/maven/internal/llm"
+)
+
+func TestPickLLMRouterOff(t *testing.T) {
+	if r := pickLLMRouter(false, llm.New("http://127.0.0.1:1", time.Second)); r != nil {
+		t.Error("flag off should give no LLM router")
+	}
+}
+
+// The operator can turn the flag on without an LLM phraser configured. That must
+// leave the classifier running, not panic.
+func TestPickLLMRouterOnWithoutClient(t *testing.T) {
+	if r := pickLLMRouter(true, nil); r != nil {
+		t.Error("no llama-server should give no LLM router")
+	}
+}
+
+func TestPickLLMRouterOn(t *testing.T) {
+	if r := pickLLMRouter(true, llm.New("http://127.0.0.1:1", time.Second)); r == nil {
+		t.Error("flag on with a client should give an LLM router")
+	}
+}

cmd/mavend/voice.go

@@ -199,8 +199,6 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	if lp, ok := phr.(*phraser.LLMPhraser); ok {
 		llmClient = llm.New(lp.BaseURL(), 60*time.Second)
 	}
-	// LLM router disabled — the classifier handles routing reliably.
-
 	// ----- router (the cascade; floor examples seed the classifier) -----
 	// The act matcher's allowlist is exactly the enabled tool names — the
 	// router only matches acts the executor can run (one source of truth).
@@ -208,7 +206,11 @@ func wireVoice(cfg *config.Config, coreAPI ipc.CoreAPI, phr phraser.Phraser, mem
 	if threshold <= 0 {
 		threshold = config.DefaultRouterThreshold
 	}
-	rtr := buildRouter(emb, matcher, threshold, nil) // LLM router disabled
+	// Both routing paths are weak on held-out utterances — the classifier gets
+	// 36.8% of intents right, the resident model 50.0% and much slower. Off by
+	// default (see config.VoiceConfig.LLMRouter); the classifier always stays
+	// wired as the fallback, so a model error never breaks a turn.
+	rtr := buildRouter(emb, matcher, threshold, pickLLMRouter(cfg.Voice.LLMRouter, llmClient))
 
 	// ----- sessions registry (shared with voicesink) -----
 	sessions := voice.NewSessions()
@@ -1048,6 +1050,21 @@ func (h *reactiveHandler) reply(ctx context.Context, text string, _ []string) (v
 	return voice.PushToTalkResp{ReplyText: text, ReplyAudio: audioOut}, nil
 }
 
+// pickLLMRouter returns the LLM router when the operator asked for it and there
+// is a llama-server to talk to, and nil otherwise. nil is safe: the cascade then
+// routes with the classifier, so an unusable setting costs accuracy, not turns.
+func pickLLMRouter(enabled bool, c *llm.Client) *router.LLMRouter {
+	if !enabled {
+		return nil
+	}
+	if c == nil {
+		log.Printf("voice: voice.llm_router is on but there is no llama-server to route with (the phraser is not an LLM phraser) — using the classifier instead")
+		return nil
+	}
+	log.Printf("voice: LLM router enabled")
+	return router.NewLLMRouter(c)
+}
+
 // buildRouter constructs the reactive-path router with the given embedder
 // and confidence threshold.
 //   - stage-0 grammars from DefaultActMatcher whose fn allowlist is exactly

deploy/mavend.json

@@ -40,6 +40,7 @@
       "tokenizer_path": "/opt/maven/models/embedder/tokenizer.json",
       "lib_path": "/opt/maven/lib/libonnxruntime.so"
     },
+    "llm_router": false,
     "tool_timeout": "30s",
     "tools": [
       { "name": "status",     "cmd": ["systemctl", "status"],         "scope": "homelab", "destructive": false },

internal/config/config.go

@@ -257,6 +257,20 @@ type VoiceConfig struct {
 	// Default 0.35 if unset.
 	RouterThreshold float64 `json:"router_threshold,omitempty"`
 
+	// LLMRouter — route with the resident model instead of the embedding
+	// classifier. Measured on the held-out fixture (ROUTING-EVAL-31-07-2026.md)
+	// the model gets 50.0% of intents right against the classifier's 36.8%, but
+	// it costs about 800ms per turn instead of 30ms.
+	//
+	// TODO: the default stays false until two things land.
+	//   1. The LLM router cannot refuse. LLMRouter.Route hardcodes
+	//      Confidence: 1.0, so the stage-3 clarify gate never fires and an
+	//      unclear utterance becomes a confident wrong action (Vikunja #359).
+	//   2. Extractor.Extract never runs on an LLM decision, so acts arrive with
+	//      no Fn and reminders with no Time.
+	// Turning this on today makes routing more accurate and less safe.
+	LLMRouter bool `json:"llm_router,omitempty"`
+
 	// QueryMinScore — the note-recall confidence gate. Top cosine below this
 	// ⇒ "I don't know" instead of a guess. Tuned for the ONNX embedder (0.55);
 	// the HashEmbedder floor scores lexically and may never clear it. 0.55

internal/config/config_test.go

@@ -171,6 +171,28 @@ func TestWeatherConfigNilOK(t *testing.T) {
 	}
 }
 
+func TestLLMRouterDefaultsOff(t *testing.T) {
+	p := writeConfig(t, `{"voice":{"enabled":true,"bind":"127.0.0.1:9100"}}`)
+	c, err := Load(p)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if c.Voice.LLMRouter {
+		t.Error("voice.llm_router absent should mean false")
+	}
+}
+
+func TestLLMRouterRead(t *testing.T) {
+	p := writeConfig(t, `{"voice":{"enabled":true,"bind":"127.0.0.1:9100","llm_router":true}}`)
+	c, err := Load(p)
+	if err != nil {
+		t.Fatalf("Load: %v", err)
+	}
+	if !c.Voice.LLMRouter {
+		t.Error("voice.llm_router true was not read")
+	}
+}
+
 func TestDurationRoundTrip(t *testing.T) {
 	d := Duration(15 * time.Minute)
 	b, err := d.MarshalJSON()

kill-maven.sh

@@ -7,12 +7,22 @@
 set -euo pipefail
 
 # The llama-server the phraser spawns has NO "maven" in its command line (its
-# args are `-m /path/to/LFM2.5-...gguf --port ...`), so a `llama-server.*maven`
+# args are `-m /path/to/<model>.gguf --port ...`), so a `llama-server.*maven`
 # pattern matches nothing and leaks it — the exact bug that let orphans pile up
-# and OOM the box. Match the model instead. Override MODEL if you change it.
-MODEL="${MODEL:-LFM2}"
+# and OOM the box.
+#
+# We used to match the model name, defaulting to LFM2. The deploy now runs
+# Qwen3.5-0.8B, so that default matched nothing and the server survived every
+# kill. Match any llama-server serving a .gguf instead, so swapping the model in
+# deploy/mavend.json cannot break this script again. Set MODEL to narrow it if
+# some other llama-server on this box must be left alone.
+MODEL="${MODEL:-}"
 PAT='mavend|mavsttd|mavttsd|mavweb|mavpoll|mavenclient'
-LLM="llama-server.*${MODEL}"
+if [ -n "$MODEL" ]; then
+    LLM="llama-server.*${MODEL}"
+else
+    LLM='llama-server.*\.gguf'
+fi
 
 echo "--- Sending graceful SIGTERM to Maven services ---"
 pkill -TERM -f "$PAT" || true