Compare commits
36
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
aefdff89dc | ||
|
|
f99e90f524 | ||
|
|
1dd47888a8 | ||
|
|
8b22557ca5 | ||
|
|
bd78f384a0 | ||
|
|
14bdd4dbf4 | ||
|
|
bbfa1f73f6 | ||
|
|
fed3cb9f18 | ||
|
|
9ec9119bd9 | ||
|
|
06d0064c8c | ||
|
|
a7cde52ee6 | ||
|
|
a353b62b37 | ||
|
|
bc47c2053b | ||
|
|
dc043ceb4d | ||
|
|
8bac7c56bc | ||
|
|
3b6d36f2de | ||
|
|
7f6e266d15 | ||
|
|
4d1664a328 | ||
|
|
1ff02f9763 | ||
|
|
5b5d61513f | ||
|
|
078ed17b57 | ||
|
|
0a2e59d756 | ||
|
|
2dfe6fd9c3 | ||
|
|
8ae20a9bd8 | ||
|
|
3ddcf665f0 | ||
|
|
7ae61701ad | ||
|
|
9cc1aec5ec | ||
|
|
53c098a9c8 | ||
|
|
0e42cf775f | ||
|
|
0e0c5d742f | ||
|
|
5b20bc1743 | ||
|
|
c3652d2f8e | ||
|
|
3cd7350160 | ||
|
|
a58aa5594d | ||
|
|
3a3c14fdc5 | ||
|
|
1d4f23ada3 |
+54
-1
@@ -10,7 +10,60 @@ Alle Änderungen am Projekt. Format: [Keep a Changelog](https://keepachangelog.c
|
||||
|
||||
---
|
||||
|
||||
## [Unreleased] — 2026-07-10
|
||||
## [0.2.0.6] — 2026-07-11
|
||||
|
||||
### Hinzugefügt
|
||||
|
||||
**Diagnostic + App — Projekte verstecken**
|
||||
- Neues `hidden`-Flag pro Projekt (bleibt voll nutzbar, nur aus Listen ausgeblendet — unabhängig von `status`/`archived`); `PATCH /projects/{id} {hidden}`
|
||||
- Diagnostic: 👁-Auge pro Projekt-Bubble (🙈 verstecken / 👁 dauerhaft sichtbar), Header-Toggle „Versteckte anzeigen (N)" blendet sie temporär gedimmt + „versteckt"-Badge ein — zum Ansehen/Auswählen ohne permanentes Enttarnen
|
||||
- App (`ProjectsBrowser`): versteckte standardmäßig ausgeblendet (Mama sieht sie nicht), Auge pro Zeile + Toggle spiegeln das Diagnostic-Verhalten; geteilter `hidden`-Status übers Brain
|
||||
|
||||
**Diagnostic — Token-Ersparnis durch lokales LLM**
|
||||
- `metrics.jsonl` trägt jetzt `source` (claude | local | fast-path); lokale Calls nutzen echte `usage`-Tokens vom Adapter, Fast-Path = 0 Prompt-Tokens (`by_source`-Aggregation, rückwärts-kompatibel)
|
||||
- Neue Card „Lokales LLM & Claude-Ersparnis": pro Fenster (1h/5h/24h/30d) gesparte Claude-Calls (local + fast-path) + lokale Token-Last (eigene HW, kein Quota)
|
||||
|
||||
**Spotify-Skill — von ARIA selbst geschärft**
|
||||
- Geräte-Transfer startet die Wiedergabe direkt mit (`play=true`) statt nur zu übertragen, inkl. Verifikation (`is_playing`-Check + expliziter Play-Fallback), Fuzzy-Gerätenamen und sauberen Exit-Codes; neue semantische Actions `play_on_device`/`search_and_play`/`playlist_play`/`queue_add`
|
||||
|
||||
### Behoben
|
||||
|
||||
- **Spotify-Resume (App):** nach einem Voice-Befehl blieb Spotify auf dem Handy pausiert. Statt des auf manchen Geräten (OnePlus) flakigen Audio-Focus-Nudge jetzt ein echter `KEYCODE_MEDIA_PLAY`-KeyEvent an die aktive MediaSession — gegated: nur wenn vor dem Dialog Musik lief (`isMusicActive`). Deterministisch, geräteunabhängig
|
||||
- **TTS-Zahlen:** freistehende Ganzzahlen werden jetzt tag-unabhängig ausgeschrieben („23°C" → „dreiundzwanzig Grad Celsius", „100%" → „einhundert Prozent"). Regression, seit das lokale LLM (bewusst ohne `<voice>`-Tag) leichte Turns übernahm; neuer vollständiger Zahl→Wort-Konverter (0…999999) am Ende von `clean_text_for_tts`, lange Ziffernfolgen (IDs) bleiben Ziffern
|
||||
- **„ARIA denkt" hängt:** Indikator + Abbrechen blieben stehen, obwohl der Turn laut Diagnostic fertig war. Die App räumt den kontext-scoped Indikator jetzt beim Eintreffen der Antwort selbst; die Bridge sendet zusätzlich ein `idle` für die Request-`projectId`, falls der Turn umgeroutet wurde (thinking ging mit Request-, idle mit Turn-`projectId`)
|
||||
- **Lokale Tool-Fehler:** Action-Skills, die bei Exit 0 einen Fehlschlag nur im stdout-Text melden (Spotify: „Fehler beim Übertragen", „Gerät nicht gefunden"), eskalieren jetzt generisch an Claude statt vom lokalen LLM vorgelesen zu werden (Info-Tools wie web_search ausgenommen)
|
||||
|
||||
---
|
||||
|
||||
## [0.2.0.4 – 0.2.0.5] — 2026-07-11 — Plan B: Lokales LLM („Gemini-Feeling")
|
||||
|
||||
Ein kleines, schnelles Modell (**Qwen3 8B** via llama.cpp/llama-swap auf der Gamebox-GPU) übernimmt einfache Turns in **<1 s**; alles Schwere/Technische/Werkzeug-artige reicht ein Router automatisch an **Claude** weiter. Ziel: schnelle Antworten ohne die Claude-Max-Subscription aufzugeben.
|
||||
|
||||
### Hinzugefügt
|
||||
|
||||
**Lokales LLM (Brain + Bridge + Adapter)**
|
||||
- Router (B1a): Heuristik + `<<ESCALATE>>`-Selbstabbruch entscheidet pro Turn lokal vs. Claude; schlanker System-Prompt mit demselben `IDENTITY_ANCHOR` wie Claude (Rolle hält), nur letzte 8 Turns (Speed)
|
||||
- Lokale Tool-Fast-Lane (B1b): kuratierte Tools — `web_search` (self-hosted **SearXNG**, local-only), `memory_search`, `trigger_timer`, Spotify; Eskalation bei Tool-Fehler statt Raten
|
||||
- Consumer-Kette gespiegelt zu FLUX: Brain → Bridge `/internal/local-llm` → RVS → `llm-adapter` → llama.cpp; `enable_thinking:false` (Qwen wickelte sonst die ganze Antwort in `<think>`)
|
||||
- **B0.5:** llama-swap (Hot-Swap der Modelle on-demand) + Modellauswahl-Dropdown + Live-Lade-Status (loading/ready + Download-Hinweis) in Diagnostic
|
||||
- **SearXNG** als 6. Container auf der ARIA-VM (keyless Meta-Suche, JSON-API)
|
||||
|
||||
**Quell-Badge (local / claude / fast-path)**
|
||||
- Diagnostic: immer an den ARIA-Bubbles
|
||||
- App: optionaler Schalter in den Einstellungen, pro Gerät gemerkt, default aus („ich will's, meine Mama nicht")
|
||||
|
||||
**TTS — System-Flag `speak` (ja/nein) pro Antwort**
|
||||
- Die Quelle entscheidet übers Vorlesen (Fast-Path/Steuerbefehl = stumm, ARIA-Antwort = vorlesen), robust statt des fragilen leeren `<voice></voice>`-Hacks der beim Skill-Rebuild verloren ging
|
||||
|
||||
### Behoben / Geändert
|
||||
|
||||
- **Identität (Hauptchat):** Proxy nutzt jetzt `--system-prompt` (voller Replace) statt `--append-system-prompt` — die Claude-Code-Basis-Identität leakt nicht mehr in den Hauptchat (ARIA antwortete dort als „Claude Code" bzw. deutete die Persona als Injection). Dazu `IDENTITY_SEED` (synthetischer Grounding-Turn) + Gift-Wächter (Identity-Breaks werden nie in die History persistiert, Retry+Fallback) + Cleanup-Script gegen bereits vergiftete Turns
|
||||
- **Datenschutz (kritisch):** harte Diskretions-Regel im `IDENTITY_ANCHOR` — ARIA kennt intime/private Details, gibt sie aber **NIE ungefragt** preis (nicht in Vorstellungen, „was weißt du über mich", Zusammenfassungen, Triggern); nur auf konkrete Nachfrage, knapp. Bereits ausgeplauderte Turns bereinigt. Lokales Tier eskaliert Personen-/Beziehungs-/Gedächtnisfragen an Claude (kennt das Gedächtnis + antwortet diskret)
|
||||
- **Lokale Antwort nicht in `<voice>`** wickeln (Qwen imitierte den Tag aus dem Kontext → Anzeige war leer); **generische** Tool-Fehler-Eskalation statt per-Skill-Router-Hardcode (Router muss nicht wissen, welche Skills „schwer" sind)
|
||||
|
||||
---
|
||||
|
||||
## [0.2.0.3] — 2026-07-10
|
||||
|
||||
### Hinzugefügt
|
||||
|
||||
|
||||
@@ -79,8 +79,8 @@ android {
|
||||
applicationId "com.ariacockpit"
|
||||
minSdkVersion rootProject.ext.minSdkVersion
|
||||
targetSdkVersion rootProject.ext.targetSdkVersion
|
||||
versionCode 20003
|
||||
versionName "0.2.0.3"
|
||||
versionCode 20008
|
||||
versionName "0.2.0.8"
|
||||
// Fallback fuer Libraries mit Product Flavors
|
||||
missingDimensionStrategy 'react-native-camera', 'general'
|
||||
}
|
||||
|
||||
@@ -5,7 +5,9 @@ import android.media.AudioAttributes
|
||||
import android.media.AudioFocusRequest
|
||||
import android.media.AudioManager
|
||||
import android.os.Build
|
||||
import android.os.SystemClock
|
||||
import android.util.Log
|
||||
import android.view.KeyEvent
|
||||
import com.facebook.react.bridge.Arguments
|
||||
import com.facebook.react.bridge.Promise
|
||||
import com.facebook.react.bridge.ReactApplicationContext
|
||||
@@ -183,6 +185,50 @@ class AudioFocusModule(reactContext: ReactApplicationContext) : ReactContextBase
|
||||
promise.resolve(true)
|
||||
}
|
||||
|
||||
/** Zuverlaessiger Spotify-Resume: einen echten MEDIA_PLAY-Tastendruck an die
|
||||
* aktive MediaSession schicken — exakt das Signal der Play-Taste am
|
||||
* Bluetooth-Kopfhoerer. Anders als nudgeMediaResume (Focus-Stack-Trick,
|
||||
* auf manchen OEMs/Spotify-Versionen unzuverlaessig) spricht das Spotifys
|
||||
* MediaSession DIREKT an und startet die Wiedergabe deterministisch wieder.
|
||||
*
|
||||
* Wir senden bewusst KEYCODE_MEDIA_PLAY (nicht PLAY_PAUSE) — das kann nur
|
||||
* starten, nie pausieren. Aufrufer muss also selbst gaten (nur senden wenn
|
||||
* vor dem Gespraech wirklich Musik lief, siehe isMusicActive()).
|
||||
*/
|
||||
@ReactMethod
|
||||
fun dispatchMediaPlay(promise: Promise) {
|
||||
val am = audioManager()
|
||||
if (am == null) {
|
||||
promise.resolve(false)
|
||||
return
|
||||
}
|
||||
try {
|
||||
val now = SystemClock.uptimeMillis()
|
||||
val down = KeyEvent(now, now, KeyEvent.ACTION_DOWN, KeyEvent.KEYCODE_MEDIA_PLAY, 0)
|
||||
val up = KeyEvent(now, now, KeyEvent.ACTION_UP, KeyEvent.KEYCODE_MEDIA_PLAY, 0)
|
||||
am.dispatchMediaKeyEvent(down)
|
||||
am.dispatchMediaKeyEvent(up)
|
||||
Log.i(TAG, "dispatchMediaPlay: KEYCODE_MEDIA_PLAY an aktive MediaSession gesendet")
|
||||
promise.resolve(true)
|
||||
} catch (e: Exception) {
|
||||
Log.w(TAG, "dispatchMediaPlay failed: ${e.message}")
|
||||
promise.resolve(false)
|
||||
}
|
||||
}
|
||||
|
||||
/** Ob gerade Musik/Media aktiv abgespielt wird (AudioManager.isMusicActive).
|
||||
* Der Aufrufer merkt sich das VOR dem Focus-Grab, um am Dialog-Ende zu
|
||||
* entscheiden ob ein dispatchMediaPlay-Resume ueberhaupt gewuenscht ist. */
|
||||
@ReactMethod
|
||||
fun isMusicActive(promise: Promise) {
|
||||
val am = audioManager()
|
||||
if (am == null) {
|
||||
promise.resolve(false)
|
||||
return
|
||||
}
|
||||
promise.resolve(am.isMusicActive)
|
||||
}
|
||||
|
||||
/** Den USAGE_MEDIA-Focus-Stack im System aufmischen, damit Spotify/YouTube
|
||||
* resumen wenn ein anderer Player (z.B. react-native-sound) seinen Focus
|
||||
* nicht ordnungsgemaess released hat. Strategie: kurz selbst USAGE_MEDIA
|
||||
|
||||
@@ -1,6 +1,6 @@
|
||||
{
|
||||
"name": "aria-cockpit",
|
||||
"version": "0.2.0.3",
|
||||
"version": "0.2.0.8",
|
||||
"private": true,
|
||||
"scripts": {
|
||||
"android": "react-native run-android",
|
||||
|
||||
@@ -75,6 +75,9 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
const [editing, setEditing] = useState<Project | null>(null);
|
||||
const [editName, setEditName] = useState('');
|
||||
const [editDesc, setEditDesc] = useState('');
|
||||
// Versteckte Projekte standardmaessig ausblenden; Toggle blendet sie
|
||||
// temporaer (gedimmt) ein — zum Ansehen/Auswaehlen oder Wieder-Sichtbarmachen.
|
||||
const [showHidden, setShowHidden] = useState(false);
|
||||
|
||||
// Refs damit useCallback NICHT bei jeder Re-Render des Parents neu erzeugt
|
||||
// wird (parent uebergibt oft inline-arrow-Callbacks, neue Identity jedes
|
||||
@@ -109,6 +112,17 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
return () => unsub();
|
||||
}, [visible, load]);
|
||||
|
||||
// Live-Sync: ein anderer Client (Diagnostic / andere App) hat ein Projekt
|
||||
// geaendert (verstecken/anlegen/beenden/…) → project_changed ueber RVS →
|
||||
// Liste neu laden, ohne dass Stefan manuell refreshen muss.
|
||||
useEffect(() => {
|
||||
if (!visible) return;
|
||||
const unsub = rvs.onMessage((msg: any) => {
|
||||
if (msg?.type === 'project_changed') load();
|
||||
});
|
||||
return () => unsub();
|
||||
}, [visible, load]);
|
||||
|
||||
const switchTo = useCallback((id: string) => {
|
||||
// Multi-Threading: Focus-Wechsel ist reine App-lokale UI-Entscheidung.
|
||||
// Brain wird nicht mehr benachrichtigt (kein globaler active_project mehr).
|
||||
@@ -158,6 +172,19 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
]);
|
||||
}, [load]);
|
||||
|
||||
// Nach einer Projekt-Mutation die anderen Clients (Diagnostic, weitere
|
||||
// App-Instanzen) live aktualisieren — via RVS project_changed. RVS echot
|
||||
// NICHT an den Sender zurueck, darum laden wir lokal zusaetzlich selbst.
|
||||
const broadcastProjectsChanged = useCallback(() => {
|
||||
try { rvs.send('project_changed' as any, { reason: 'app' }); } catch {}
|
||||
}, []);
|
||||
|
||||
const toggleHidden = useCallback((p: Project) => {
|
||||
brainApi.setProjectHidden(p.id, !p.hidden)
|
||||
.then(() => { broadcastProjectsChanged(); load(); })
|
||||
.catch(e => Alert.alert('Fehler', String(e?.message || e)));
|
||||
}, [load, broadcastProjectsChanged]);
|
||||
|
||||
const archiveProject = useCallback((p: Project) => {
|
||||
Alert.alert(`"${p.name}" archivieren?`,
|
||||
'Verschwindet aus der Standardliste. Über "archivierte zeigen" erreichbar.',
|
||||
@@ -176,11 +203,12 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
const renderItem = ({ item }: { item: Project }) => {
|
||||
const isActive = item.id === activeId;
|
||||
const dot = _statusDot(item.id);
|
||||
const hidden = !!item.hidden;
|
||||
return (
|
||||
<TouchableOpacity
|
||||
onPress={() => switchTo(item.id)}
|
||||
onLongPress={() => openEdit(item)}
|
||||
style={[s.row, isActive && s.rowActive]}
|
||||
style={[s.row, isActive && s.rowActive, hidden && s.rowHidden]}
|
||||
>
|
||||
<View style={{ flex: 1 }}>
|
||||
<View style={{ flexDirection: 'row', alignItems: 'center', gap: 8 }}>
|
||||
@@ -188,6 +216,7 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
<View style={{ width: 8, height: 8, borderRadius: 4, backgroundColor: dot.color }} />
|
||||
)}
|
||||
<Text style={[s.rowName, isActive && { color: '#34C759' }]}>{item.name}</Text>
|
||||
{hidden && <Text style={s.hiddenBadge}>versteckt</Text>}
|
||||
{item.status === 'ended' && <Text style={s.statusBadge}>beendet</Text>}
|
||||
{isActive && <Text style={s.activeBadge}>✓ FOCUS</Text>}
|
||||
</View>
|
||||
@@ -199,10 +228,22 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
{dot.label ? ` · ${dot.label}` : ''}
|
||||
</Text>
|
||||
</View>
|
||||
{/* Auge: verstecken (🙈) / wieder sichtbar (👁). Eigener Touch, damit
|
||||
der Tap NICHT das Projekt wechselt. */}
|
||||
<TouchableOpacity
|
||||
onPress={() => toggleHidden(item)}
|
||||
hitSlop={{ top: 10, bottom: 10, left: 10, right: 10 }}
|
||||
style={s.eyeBtn}
|
||||
>
|
||||
<Text style={s.eyeIcon}>{hidden ? '👁' : '🙈'}</Text>
|
||||
</TouchableOpacity>
|
||||
</TouchableOpacity>
|
||||
);
|
||||
};
|
||||
|
||||
const hiddenCount = projects.filter(p => p.hidden).length;
|
||||
const visibleProjects = showHidden ? projects : projects.filter(p => !p.hidden);
|
||||
|
||||
const body = (
|
||||
<View style={{ flex: 1, backgroundColor: '#0A0A14' }}>
|
||||
{/* Header */}
|
||||
@@ -243,6 +284,17 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
);
|
||||
})()}
|
||||
|
||||
{/* Versteckte-Toggle — nur wenn es welche gibt (oder gerade eingeblendet) */}
|
||||
{(hiddenCount > 0 || showHidden) && (
|
||||
<TouchableOpacity onPress={() => setShowHidden(v => !v)} style={s.hiddenToggle}>
|
||||
<Text style={s.hiddenToggleText}>
|
||||
{showHidden
|
||||
? `🙈 Versteckte ausblenden${hiddenCount ? ` (${hiddenCount})` : ''}`
|
||||
: `👁 Versteckte anzeigen${hiddenCount ? ` (${hiddenCount})` : ''}`}
|
||||
</Text>
|
||||
</TouchableOpacity>
|
||||
)}
|
||||
|
||||
{loading ? (
|
||||
<View style={{ padding: 24, alignItems: 'center' }}>
|
||||
<ActivityIndicator color="#0096FF" />
|
||||
@@ -251,14 +303,21 @@ export const ProjectsBrowser: React.FC<Props> = ({ visible = true, onClose, onAc
|
||||
<Text style={s.errorText}>⚠ {err}</Text>
|
||||
) : (
|
||||
<FlatList
|
||||
data={projects}
|
||||
data={visibleProjects}
|
||||
keyExtractor={p => p.id}
|
||||
renderItem={renderItem}
|
||||
ListEmptyComponent={
|
||||
<Text style={s.emptyText}>
|
||||
Noch keine Projekte. Tipp + Neu oder sag zu ARIA:{'\n'}
|
||||
„Lass uns ein Projekt 'XY' anlegen".
|
||||
</Text>
|
||||
projects.length > 0 ? (
|
||||
<Text style={s.emptyText}>
|
||||
Alle {hiddenCount} Projekte sind versteckt.{'\n'}
|
||||
Tipp „👁 Versteckte anzeigen".
|
||||
</Text>
|
||||
) : (
|
||||
<Text style={s.emptyText}>
|
||||
Noch keine Projekte. Tipp + Neu oder sag zu ARIA:{'\n'}
|
||||
„Lass uns ein Projekt 'XY' anlegen".
|
||||
</Text>
|
||||
)
|
||||
}
|
||||
/>
|
||||
)}
|
||||
@@ -365,6 +424,8 @@ const s = StyleSheet.create({
|
||||
headerBtnText: { color: '#0096FF', fontSize: 18, fontWeight: '600' },
|
||||
headerTitle: { flex: 1, textAlign: 'center', color: '#E0E0F0', fontSize: 18, fontWeight: '700' },
|
||||
row: {
|
||||
flexDirection: 'row',
|
||||
alignItems: 'center',
|
||||
paddingHorizontal: 16,
|
||||
paddingVertical: 12,
|
||||
borderBottomWidth: 1,
|
||||
@@ -375,6 +436,18 @@ const s = StyleSheet.create({
|
||||
borderLeftWidth: 3,
|
||||
borderLeftColor: '#34C759',
|
||||
},
|
||||
rowHidden: { opacity: 0.55 },
|
||||
eyeBtn: { paddingHorizontal: 8, paddingVertical: 6, marginLeft: 6 },
|
||||
eyeIcon: { fontSize: 18 },
|
||||
hiddenBadge: { color: '#B392F0', fontSize: 10, fontWeight: '700',
|
||||
backgroundColor: 'rgba(179,146,240,0.15)', paddingHorizontal: 6,
|
||||
paddingVertical: 2, borderRadius: 4 },
|
||||
hiddenToggle: {
|
||||
paddingHorizontal: 16, paddingVertical: 10,
|
||||
borderBottomWidth: 1, borderColor: '#1E1E2E',
|
||||
backgroundColor: '#0D0D18',
|
||||
},
|
||||
hiddenToggleText: { color: '#B392F0', fontSize: 12, fontWeight: '600' },
|
||||
rowName: { color: '#E0E0F0', fontSize: 16, fontWeight: '600' },
|
||||
rowDesc: { color: '#8888AA', fontSize: 13, marginTop: 4 },
|
||||
rowMeta: { color: '#555570', fontSize: 11, marginTop: 4 },
|
||||
|
||||
@@ -49,7 +49,7 @@ import VoiceButton from '../components/VoiceButton';
|
||||
import FileUpload, { FileData } from '../components/FileUpload';
|
||||
import CameraUpload, { PhotoData } from '../components/CameraUpload';
|
||||
import MessageText from '../components/MessageText';
|
||||
import { loadConvWindowMs, loadTtsSpeed, TTS_SPEED_DEFAULT } from '../services/audio';
|
||||
import { loadConvWindowMs, loadTtsSpeed, TTS_SPEED_DEFAULT, loadSttEndpointMs } from '../services/audio';
|
||||
import Geolocation from '@react-native-community/geolocation';
|
||||
|
||||
// --- Typen ---
|
||||
@@ -73,6 +73,9 @@ interface ChatMessage {
|
||||
/** Projekt-Zuordnung — leer = Hauptchat. Wird genutzt um Bubbles zu
|
||||
* Projekt-Bloecken zu gruppieren (auf/einklappbar). */
|
||||
projectId?: string;
|
||||
/** Welcher Backend die Antwort erzeugt hat: 'local' | 'claude' | 'fast-path'.
|
||||
* Fuer den optionalen Quell-Badge (Einstellung aria_show_source, default aus). */
|
||||
answeredBy?: string;
|
||||
/** Bridge-Message-ID zur Zuordnung von TTS-Audio */
|
||||
messageId?: string;
|
||||
/** Lokaler Pfad zur gecachten TTS-Audio-Datei (file://...) */
|
||||
@@ -324,6 +327,8 @@ const ChatScreen: React.FC = () => {
|
||||
// App soll sie standardmaessig NICHT anzeigen — Stefan sieht sonst
|
||||
// jeden Hint mit. Toggle in Settings.
|
||||
const [showSystemHints, setShowSystemHints] = useState(false);
|
||||
// Quell-Badge (local/claude/fast-path) an ARIA-Bubbles — pro Geraet, default AUS.
|
||||
const [showSource, setShowSource] = useState(false);
|
||||
// Gerätelokale XTTS-Voice-Wahl (bevorzugt gegenueber dem globalen Default)
|
||||
const localXttsVoiceRef = useRef<string>('');
|
||||
// Geraetelokale TTS-Wiedergabegeschwindigkeit (speed-Param an F5-TTS)
|
||||
@@ -546,6 +551,8 @@ const ChatScreen: React.FC = () => {
|
||||
setGpsEnabled(gps === 'true');
|
||||
const hints = await AsyncStorage.getItem('aria_show_hints');
|
||||
setShowSystemHints(hints === 'true'); // default false
|
||||
const src = await AsyncStorage.getItem('aria_show_source');
|
||||
setShowSource(src === 'true'); // default false (Mama sieht keine Badges)
|
||||
};
|
||||
loadSettings();
|
||||
const interval = setInterval(loadSettings, 2000);
|
||||
@@ -555,6 +562,14 @@ const ChatScreen: React.FC = () => {
|
||||
// Wake Word: einmalig laden + Porcupine vorbereiten (wenn Access Key gesetzt)
|
||||
useEffect(() => {
|
||||
wakeWordService.loadFromStorage().catch(() => {});
|
||||
// Mikro-Release-Hook: vor jedem Re-Arm eine laufende (passive) Streaming-
|
||||
// Aufnahme canceln, sonst haelt sie das Mikro und OpenWakeWord.start()
|
||||
// schlaegt fehl → Ohr bleibt ausgegraut (state=off).
|
||||
wakeWordService.setMicReleaseHook(async () => {
|
||||
if (audioService.isStreamingRecording()) {
|
||||
await audioService.cancelStreamingRecording('rearm-mic-free');
|
||||
}
|
||||
});
|
||||
const unsub = wakeWordService.onStateChange((s) => {
|
||||
setWakeWordState(s);
|
||||
setWakeWordActive(s !== 'off');
|
||||
@@ -815,6 +830,7 @@ const ChatScreen: React.FC = () => {
|
||||
attachments: attachments.length ? attachments : undefined,
|
||||
backupTs: typeof m.ts === 'number' ? m.ts : undefined,
|
||||
projectId: typeof m.project_id === 'string' ? m.project_id : '',
|
||||
answeredBy: typeof m.answeredBy === 'string' ? m.answeredBy : '',
|
||||
...(cmid && { clientMsgId: cmid }),
|
||||
// Server-Bubble = vom Brain verarbeitet → 'delivered' (✓✓)
|
||||
...(role === 'user' && cmid && { deliveryStatus: 'delivered' as const }),
|
||||
@@ -1151,6 +1167,7 @@ const ChatScreen: React.FC = () => {
|
||||
messageId: (message.payload.messageId as string) || undefined,
|
||||
backupTs: (message.payload.backupTs as number) || undefined,
|
||||
projectId: ((message.payload as any).projectId as string) || '',
|
||||
answeredBy: ((message.payload as any).answeredBy as string) || '',
|
||||
};
|
||||
// ARIA hat geantwortet → alle User-Bubbles davor als 'delivered'
|
||||
// markieren (WhatsApp-Doppelhaken ✓✓). Brain hat sie verarbeitet.
|
||||
@@ -1163,6 +1180,21 @@ const ChatScreen: React.FC = () => {
|
||||
});
|
||||
// ARIA hat geantwortet → Watchdog clearen, falls noch armiert
|
||||
clearStuckWatchdog();
|
||||
// Thinking-Indikator fuer DIESEN Kontext raeumen. Die Antwort ist der
|
||||
// definitive Beweis, dass der Turn fertig ist — unabhaengig davon, ob
|
||||
// das agent_activity 'idle' der Bridge ankam (es kann mit abweichender
|
||||
// projectId gesendet oder wegdedupt worden sein → "ARIA denkt" blieb
|
||||
// sonst haengen). Zusaetzlich global raeumen, falls der sichtbare
|
||||
// Kontext ueber den Legacy-Indikator lief.
|
||||
{
|
||||
const ansPid = ((message.payload as any).projectId as string) || '';
|
||||
setAgentActivityByCtx(prev =>
|
||||
prev[ansPid] && prev[ansPid].activity !== 'idle'
|
||||
? { ...prev, [ansPid]: { activity: 'idle', tool: '' } }
|
||||
: prev);
|
||||
setAgentActivity(cur =>
|
||||
cur.activity === 'idle' ? cur : { activity: 'idle', tool: '' });
|
||||
}
|
||||
// ALLE noch laufenden ACK-Timer clearen — Bridge hat unsere Messages
|
||||
// ja offensichtlich verarbeitet (sonst keine ARIA-Antwort). Wenn
|
||||
// ein ACK aus Netzgruenden verloren ging, soll der Retry nicht
|
||||
@@ -1441,7 +1473,7 @@ const ChatScreen: React.FC = () => {
|
||||
interrupted: wasInterrupted,
|
||||
location: location || null,
|
||||
noSpeechTimeoutMs: windowMs,
|
||||
endpointMs: 1500,
|
||||
endpointMs: await loadSttEndpointMs(),
|
||||
hardCapMs: 60000,
|
||||
projectId: focusedProjectIdRef.current,
|
||||
});
|
||||
@@ -1537,7 +1569,7 @@ const ChatScreen: React.FC = () => {
|
||||
interrupted: true, // Barge-In → Brain weiss "User hat unterbrochen"
|
||||
location: location || null,
|
||||
noSpeechTimeoutMs: windowMs,
|
||||
endpointMs: 1500,
|
||||
endpointMs: await loadSttEndpointMs(),
|
||||
hardCapMs: 60000,
|
||||
projectId: focusedProjectIdRef.current,
|
||||
});
|
||||
@@ -1594,7 +1626,7 @@ const ChatScreen: React.FC = () => {
|
||||
interrupted: false,
|
||||
location: location || null,
|
||||
noSpeechTimeoutMs: Math.min(passiveMs, 30000),
|
||||
endpointMs: 1500,
|
||||
endpointMs: await loadSttEndpointMs(),
|
||||
hardCapMs: Math.max(passiveMs + 5000, 35000),
|
||||
projectId: focusedProjectIdRef.current,
|
||||
});
|
||||
@@ -2004,7 +2036,7 @@ const ChatScreen: React.FC = () => {
|
||||
// Stoppen). Hard-Cap 5 Minuten als Notbremse — danach killt Whisper
|
||||
// die Session auch app-seitig haben wir +2s Toleranz.
|
||||
noSpeechTimeoutMs: 0,
|
||||
endpointMs: 1500,
|
||||
endpointMs: await loadSttEndpointMs(),
|
||||
hardCapMs: 300000,
|
||||
projectId: focusedProjectIdRef.current,
|
||||
});
|
||||
@@ -2411,6 +2443,16 @@ const ChatScreen: React.FC = () => {
|
||||
) : null}
|
||||
<View style={styles.statusRow}>
|
||||
<Text style={styles.timestamp}>{time}</Text>
|
||||
{!isUser && showSource && item.answeredBy ? (() => {
|
||||
const ab = item.answeredBy as string;
|
||||
const map: { [k: string]: { t: string; c: string } } = {
|
||||
local: { t: '⚡ lokal', c: '#34C759' },
|
||||
claude: { t: 'Claude', c: '#0096FF' },
|
||||
'fast-path': { t: '⚡ fast', c: '#AF7BFF' },
|
||||
};
|
||||
const b = map[ab] || { t: ab, c: '#8A8AA0' };
|
||||
return <Text style={{ fontSize: 9, fontWeight: 'bold', color: b.c, marginLeft: 6 }}>{b.t}</Text>;
|
||||
})() : null}
|
||||
{item.text.length > 0 ? (
|
||||
<TouchableOpacity
|
||||
hitSlop={{top:6,bottom:6,left:6,right:6}}
|
||||
|
||||
@@ -175,6 +175,7 @@ const SettingsScreen: React.FC = () => {
|
||||
const [bgGpsEnabled, setBgGpsEnabled] = useState(false);
|
||||
const [backgroundMode, setBackgroundMode] = useState(true); // Default an
|
||||
const [showSystemHints, setShowSystemHints] = useState(false); // Default aus
|
||||
const [showSource, setShowSource] = useState(false); // Quell-Badge, Default aus
|
||||
const [scannerVisible, setScannerVisible] = useState(false);
|
||||
const [logTab, setLogTab] = useState<LogTab>('live');
|
||||
const [logs, setLogs] = useState<LogEntry[]>([]);
|
||||
@@ -261,6 +262,9 @@ const SettingsScreen: React.FC = () => {
|
||||
// Default ist aus — nur explicit 'true' aktiviert
|
||||
setShowSystemHints(saved === 'true');
|
||||
});
|
||||
AsyncStorage.getItem('aria_show_source').then(saved => {
|
||||
setShowSource(saved === 'true'); // Default aus
|
||||
});
|
||||
// gpsTrackingService status syncen + auf Aenderungen lauschen
|
||||
setGpsTracking(gpsTrackingService.isActive());
|
||||
const offGps = gpsTrackingService.onChange(setGpsTracking);
|
||||
@@ -857,6 +861,11 @@ const SettingsScreen: React.FC = () => {
|
||||
AsyncStorage.setItem('aria_show_hints', String(value)).catch(() => {});
|
||||
}, []);
|
||||
|
||||
const handleShowSourceToggle = useCallback((value: boolean) => {
|
||||
setShowSource(value);
|
||||
AsyncStorage.setItem('aria_show_source', String(value)).catch(() => {});
|
||||
}, []);
|
||||
|
||||
// --- XTTS Voice ---
|
||||
|
||||
const selectVoice = useCallback((voiceName: string) => {
|
||||
@@ -1580,6 +1589,22 @@ const SettingsScreen: React.FC = () => {
|
||||
thumbColor={showSystemHints ? '#FFFFFF' : '#666680'}
|
||||
/>
|
||||
</View>
|
||||
<View style={styles.toggleRow}>
|
||||
<View style={styles.toggleInfo}>
|
||||
<Text style={styles.toggleLabel}>Antwort-Quelle anzeigen</Text>
|
||||
<Text style={styles.toggleHint}>
|
||||
Kleiner Badge an ARIAs Bubbles: ob die Antwort vom schnellen
|
||||
lokalen Modell, von Claude oder per Direkt-Befehl kam. Nur fuer
|
||||
dich interessant (Technik) — standardmaessig aus.
|
||||
</Text>
|
||||
</View>
|
||||
<Switch
|
||||
value={showSource}
|
||||
onValueChange={handleShowSourceToggle}
|
||||
trackColor={{ false: '#2A2A3E', true: '#0096FF' }}
|
||||
thumbColor={showSource ? '#FFFFFF' : '#666680'}
|
||||
/>
|
||||
</View>
|
||||
</View>
|
||||
|
||||
{/* === Hintergrund-Modus === */}
|
||||
|
||||
@@ -44,6 +44,11 @@ const { AudioFocus, PcmStreamPlayer, PcmStreamRecorder } = NativeModules as {
|
||||
release: () => Promise<boolean>;
|
||||
kickReleaseMedia: () => Promise<boolean>;
|
||||
getMode?: () => Promise<number>;
|
||||
// Zuverlaessiger Spotify-Resume via echtem MEDIA_PLAY-KeyEvent (statt
|
||||
// Focus-Stack-Nudge). isMusicActive() zum Gaten: nur resumen wenn vor
|
||||
// dem Gespraech wirklich Musik lief.
|
||||
dispatchMediaPlay?: () => Promise<boolean>;
|
||||
isMusicActive?: () => Promise<boolean>;
|
||||
};
|
||||
PcmStreamPlayer?: {
|
||||
start: (sampleRate: number, channels: number, prerollSeconds: number) => Promise<boolean>;
|
||||
@@ -146,6 +151,26 @@ export const CONV_WINDOW_MIN_SEC = 3.0;
|
||||
export const CONV_WINDOW_MAX_SEC = 20.0;
|
||||
export const CONV_WINDOW_STORAGE_KEY = 'aria_conv_window_sec';
|
||||
|
||||
// STT-Endpoint (ms Stille bis "fertig gesprochen"). Zu kurz = schneidet mitten
|
||||
// im Satz ab, besonders im Auto wo man mit Pausen spricht (Reproduktion: die
|
||||
// 11.8s-Frage wurde bei "…ohne dass ein" gekappt). 1500 war zu aggressiv;
|
||||
// 2400 default, im Auto ggf. hoeher. Konfigurierbar in den Settings.
|
||||
export const STT_ENDPOINT_DEFAULT_MS = 2400;
|
||||
export const STT_ENDPOINT_MIN_MS = 1000;
|
||||
export const STT_ENDPOINT_MAX_MS = 4000;
|
||||
export const STT_ENDPOINT_STORAGE_KEY = 'aria_stt_endpoint_ms';
|
||||
|
||||
export async function loadSttEndpointMs(): Promise<number> {
|
||||
try {
|
||||
const raw = await AsyncStorage.getItem(STT_ENDPOINT_STORAGE_KEY);
|
||||
if (raw != null) {
|
||||
const n = parseInt(raw, 10);
|
||||
if (isFinite(n) && n >= STT_ENDPOINT_MIN_MS && n <= STT_ENDPOINT_MAX_MS) return n;
|
||||
}
|
||||
} catch {}
|
||||
return STT_ENDPOINT_DEFAULT_MS;
|
||||
}
|
||||
|
||||
// TTS-Wiedergabegeschwindigkeit — wird pro Geraet gespeichert und an die
|
||||
// Bridge mitgegeben (speed-Param im F5-TTS infer()). 1.0 = normal.
|
||||
export const TTS_SPEED_DEFAULT = 1.0;
|
||||
@@ -275,6 +300,13 @@ class AudioService {
|
||||
// damit Spotify nicht in Render-Pausen oder zwischen Antworten zurueckkehrt.
|
||||
private _conversationFocusActive: boolean = false;
|
||||
|
||||
// Lief unmittelbar VOR dem Focus-Grab (Wake-Word/Aufnahme) Musik? Wird beim
|
||||
// Betreten des Dialogs gemerkt (latch: nur auf true), damit wir am Dialog-Ende
|
||||
// NUR dann Spotify per MEDIA_PLAY-KeyEvent zuverlaessig resumen, wenn vorher
|
||||
// wirklich etwas lief. Verhindert, dass wir bei Stille versehentlich Musik
|
||||
// starten. Wird nach dem Resume-Dispatch wieder auf false gesetzt.
|
||||
private _mediaWasActiveAtAcquire: boolean = false;
|
||||
|
||||
// VAD State
|
||||
private vadEnabled: boolean = false;
|
||||
private lastSpeechTime: number = 0;
|
||||
@@ -450,15 +482,29 @@ class AudioService {
|
||||
import('./logger').then(m => m.reportAppDebug('audio.focus',
|
||||
'AudioFocus.release() now')).catch(()=>{});
|
||||
AudioFocus?.release().catch(() => {});
|
||||
// Spotify-Resume-Trigger: nach Abandon den USAGE_MEDIA-Focus-Stack
|
||||
// mit kurzem TRANSIENT-Nudge aufmischen. Spotify resumed sonst bei
|
||||
// manchen Versionen / Geraeten nicht zuverlaessig nach Auto-Loss.
|
||||
// 50ms Delay damit das Abandon erst durch ist.
|
||||
setTimeout(() => {
|
||||
// Spotify-Resume: NUR wenn vor dem Gespraech wirklich Musik lief. Dann
|
||||
// einen echten MEDIA_PLAY-KeyEvent an die aktive MediaSession schicken
|
||||
// (wie die Play-Taste am Kopfhoerer) — das resumt Spotify zuverlaessig,
|
||||
// im Gegensatz zum flakigen Focus-Stack-Nudge, der auf manchen Geraeten
|
||||
// (OnePlus) nach Auto-Loss nicht griff. 120ms Delay, damit das Abandon
|
||||
// sicher durch ist, bevor der Play-Key kommt.
|
||||
const shouldResume = this._mediaWasActiveAtAcquire;
|
||||
this._mediaWasActiveAtAcquire = false;
|
||||
if (shouldResume) {
|
||||
setTimeout(() => {
|
||||
import('./logger').then(m => m.reportAppDebug('audio.focus',
|
||||
'dispatchMediaPlay() now (Musik lief vor Dialog → Resume)')).catch(()=>{});
|
||||
if (AudioFocus?.dispatchMediaPlay) {
|
||||
AudioFocus.dispatchMediaPlay().catch(() => {});
|
||||
} else {
|
||||
// Fallback fuer alte Native-Builds ohne dispatchMediaPlay
|
||||
AudioFocus?.nudgeMediaResume().catch(() => {});
|
||||
}
|
||||
}, 120);
|
||||
} else {
|
||||
import('./logger').then(m => m.reportAppDebug('audio.focus',
|
||||
'nudgeMediaResume() now (50ms after release)')).catch(()=>{});
|
||||
AudioFocus?.nudgeMediaResume().catch(() => {});
|
||||
}, 50);
|
||||
'kein Resume (vor Dialog lief keine Musik)')).catch(()=>{});
|
||||
}
|
||||
}, this.FOCUS_RELEASE_DELAY_MS);
|
||||
}
|
||||
|
||||
@@ -469,6 +515,20 @@ class AudioService {
|
||||
}
|
||||
}
|
||||
|
||||
/** Merkt sich (latch: nur auf true), ob GERADE Musik laeuft — VOR einem
|
||||
* Focus-Grab aufrufen. Am Dialog-Ende entscheidet die Flag, ob wir Spotify
|
||||
* aktiv per MEDIA_PLAY resumen. Awaitet bewusst isMusicActive bevor der
|
||||
* Focus-Request die Wiedergabe pausiert (sonst laese man schon 'false'). */
|
||||
private async _captureMediaActive(): Promise<void> {
|
||||
try {
|
||||
const active = await AudioFocus?.isMusicActive?.();
|
||||
if (active) {
|
||||
this._mediaWasActiveAtAcquire = true;
|
||||
console.log('[Audio] Musik lief vor Focus-Grab → Resume am Dialog-Ende gemerkt');
|
||||
}
|
||||
} catch {}
|
||||
}
|
||||
|
||||
/** Conversation-Mode beginnt → AudioFocus dauerhaft halten (Spotify bleibt
|
||||
* pausiert). Idempotent: mehrfaches Aufrufen ist sicher. */
|
||||
acquireConversationFocus(): void {
|
||||
@@ -476,7 +536,11 @@ class AudioService {
|
||||
this._conversationFocusActive = true;
|
||||
this._cancelDeferredFocusRelease();
|
||||
console.log('[Audio] Conversation-Focus aktiv (Spotify bleibt gepaust)');
|
||||
AudioFocus?.requestDuck().catch(() => {});
|
||||
// Erst pruefen ob Musik laeuft, DANN ducken (requestDuck wuerde sie sonst
|
||||
// schon pausieren bevor wir es messen koennen).
|
||||
this._captureMediaActive().finally(() => {
|
||||
AudioFocus?.requestDuck().catch(() => {});
|
||||
});
|
||||
}
|
||||
|
||||
/** Conversation-Mode endet → Focus darf wieder freigegeben werden
|
||||
@@ -493,6 +557,8 @@ class AudioService {
|
||||
haltAllPlayback(reason: string = ''): void {
|
||||
console.log('[Audio] haltAllPlayback: %s', reason || '(no reason)');
|
||||
this._conversationFocusActive = false;
|
||||
// Barge-In → User spricht gleich weiter, Spotify NICHT resumen.
|
||||
this._mediaWasActiveAtAcquire = false;
|
||||
this.stopPlayback();
|
||||
}
|
||||
|
||||
@@ -503,6 +569,8 @@ class AudioService {
|
||||
pauseForCall(reason: string = ''): void {
|
||||
console.log('[Audio] pauseForCall: %s', reason || '(no reason)');
|
||||
this._conversationFocusActive = false;
|
||||
// Anruf → Spotify bleibt aus, kein Auto-Resume beim spaeteren Release.
|
||||
this._mediaWasActiveAtAcquire = false;
|
||||
this._pausedForCall = true;
|
||||
// Queue + isPlaying ruecksetzen — sonst klemmt der naechste Play-Button
|
||||
// (playAudio sieht isPlaying=true und ruft _playNext nicht mehr auf).
|
||||
@@ -796,6 +864,8 @@ class AudioService {
|
||||
this.setState('recording');
|
||||
|
||||
// Andere Apps waehrend der Aufnahme pausieren (Musik, Videos etc.)
|
||||
// Vorher merken ob Musik lief, damit wir sie danach resumen koennen.
|
||||
await this._captureMediaActive();
|
||||
this._cancelDeferredFocusRelease();
|
||||
AudioFocus?.requestExclusive().catch(() => {});
|
||||
|
||||
@@ -1043,6 +1113,8 @@ class AudioService {
|
||||
}
|
||||
|
||||
// AudioFocus exklusiv — gleiche Semantik wie beim Legacy-Pfad.
|
||||
// Vorher merken ob Musik lief (fuer Resume am Dialog-Ende).
|
||||
await this._captureMediaActive();
|
||||
this._cancelDeferredFocusRelease();
|
||||
AudioFocus?.requestExclusive().catch(() => {});
|
||||
|
||||
|
||||
@@ -157,6 +157,7 @@ export interface Project {
|
||||
name: string;
|
||||
description: string;
|
||||
status: 'active' | 'ended' | 'archived';
|
||||
hidden?: boolean; // aus Listen ausgeblendet (bleibt nutzbar)
|
||||
created_at: number;
|
||||
updated_at: number;
|
||||
last_activity_at: number;
|
||||
@@ -593,14 +594,22 @@ export const brainApi = {
|
||||
});
|
||||
},
|
||||
|
||||
/** Projekt-Metadaten patchen (name / description). */
|
||||
updateProject(projectId: string, patch: Partial<Pick<Project, 'name' | 'description'>>): Promise<Project> {
|
||||
/** Projekt-Metadaten patchen (name / description / hidden). */
|
||||
updateProject(projectId: string, patch: Partial<Pick<Project, 'name' | 'description' | 'hidden'>>): Promise<Project> {
|
||||
return _send(`/projects/${encodeURIComponent(projectId)}`, {
|
||||
method: 'PATCH',
|
||||
body: patch,
|
||||
});
|
||||
},
|
||||
|
||||
/** Projekt verstecken / wieder sichtbar machen (bleibt voll nutzbar). */
|
||||
setProjectHidden(projectId: string, hidden: boolean): Promise<Project> {
|
||||
return _send(`/projects/${encodeURIComponent(projectId)}`, {
|
||||
method: 'PATCH',
|
||||
body: { hidden },
|
||||
});
|
||||
},
|
||||
|
||||
/** Queue-Status: pro Kontext (project_id oder __main__ fuer Hauptchat)
|
||||
* ob gerade ein Request in Verarbeitung ist + wieviele in der Queue warten.
|
||||
* Wird fuer Status-Dots im Drawer periodisch gepollt. */
|
||||
|
||||
@@ -132,6 +132,12 @@ class WakeWordService {
|
||||
* hier eine streaming-Aufnahme OHNE User-Bubble (passiv lauschen). */
|
||||
private passiveListenCallbacks: PassiveListenCallback[] = [];
|
||||
|
||||
/** Hook, der das Mikro freigibt (laufende Streaming-Aufnahme canceln) BEVOR
|
||||
* wir OpenWakeWord.start() rufen. Ohne das haelt die passive/conversing
|
||||
* Aufnahme das Mikro noch, start() schlaegt fehl → Ohr bleibt ausgegraut
|
||||
* (state=off). ChatScreen registriert den Hook mit audioService.cancel…. */
|
||||
private micReleaseHook: (() => Promise<void>) | null = null;
|
||||
|
||||
private keyword: WakeKeyword = DEFAULT_KEYWORD;
|
||||
private nativeReady: boolean = false;
|
||||
private initInProgress: Promise<boolean> | null = null;
|
||||
@@ -149,6 +155,19 @@ class WakeWordService {
|
||||
}
|
||||
}
|
||||
|
||||
/** ChatScreen registriert hier einen Hook, der eine laufende Streaming-
|
||||
* Aufnahme cancelt (Mikro freigeben) — wird vor jedem Re-Arm gerufen. */
|
||||
setMicReleaseHook(fn: (() => Promise<void>) | null): void {
|
||||
this.micReleaseHook = fn;
|
||||
}
|
||||
|
||||
private async _freeMic(): Promise<void> {
|
||||
if (!this.micReleaseHook) return;
|
||||
try { await this.micReleaseHook(); } catch (e) {
|
||||
console.warn('[WakeWord] micReleaseHook err:', e);
|
||||
}
|
||||
}
|
||||
|
||||
/** Settings-Wechsel: anderes Wake-Word. Re-Init des Native-Moduls. */
|
||||
async configure(keyword: string): Promise<boolean> {
|
||||
const next: WakeKeyword = (WAKE_KEYWORDS as readonly string[]).includes(keyword)
|
||||
@@ -454,6 +473,7 @@ class WakeWordService {
|
||||
// als state extra zu fragen, garantiert dass nach diesem Pfad
|
||||
// Native auch wirklich an ist falls es out-of-band gestoppt wurde.
|
||||
try {
|
||||
await this._freeMic(); // Streaming-Aufnahme canceln → Mikro frei
|
||||
await OpenWakeWord.start();
|
||||
console.log('[WakeWord] Konversation zu Ende — zurueck zu armed (wasBarge=%s)', wasBarge);
|
||||
import('./logger').then(m => m.reportAppDebug('wake.end',
|
||||
@@ -520,6 +540,7 @@ class WakeWordService {
|
||||
// timeout oder manual → Wake-Word reaktivieren, armed-State.
|
||||
if (this.nativeReady && OpenWakeWord) {
|
||||
try {
|
||||
await this._freeMic(); // passive Streaming-Aufnahme canceln → Mikro frei
|
||||
await OpenWakeWord.start();
|
||||
console.log('[WakeWord] zurueck zu armed nach passive-listen');
|
||||
ToastAndroid.show(`Lausche wieder auf "${KEYWORD_LABELS[this.keyword]}"`, ToastAndroid.SHORT);
|
||||
@@ -564,6 +585,7 @@ class WakeWordService {
|
||||
this.lastTriggerAt = 0;
|
||||
if (this.nativeReady && OpenWakeWord) {
|
||||
try {
|
||||
await this._freeMic(); // ggf. laufende Aufnahme canceln → Mikro frei
|
||||
await OpenWakeWord.start();
|
||||
ToastAndroid.show('Hintergrund-Trigger verworfen — lausche wieder', ToastAndroid.SHORT);
|
||||
this.setState('armed');
|
||||
|
||||
+301
-50
@@ -21,14 +21,16 @@ import logging
|
||||
import os
|
||||
import re
|
||||
import urllib.error
|
||||
import urllib.parse
|
||||
import urllib.request
|
||||
from typing import Optional
|
||||
|
||||
from conversation import Conversation, Turn
|
||||
from memory import Embedder, VectorStore, MemoryPoint
|
||||
from prompts import build_system_prompt, IDENTITY_SEED, IDENTITY_ANCHOR
|
||||
from prompts import build_system_prompt, IDENTITY_SEED, IDENTITY_ANCHOR, looks_like_identity_break
|
||||
from proxy_client import ProxyClient, Message as ProxyMessage
|
||||
import router as router_mod
|
||||
import metrics
|
||||
from local_llm import local_llm_chat
|
||||
import skills as skills_mod
|
||||
import triggers as triggers_mod
|
||||
@@ -37,6 +39,8 @@ import oauth as oauth_mod
|
||||
import projects as projects_mod
|
||||
|
||||
BRIDGE_URL = os.environ.get("BRIDGE_URL", "http://aria-bridge:8090")
|
||||
# SearXNG (self-hosted Meta-Suche) — Backend fuers web_search-Tool (B1b).
|
||||
SEARXNG_URL = os.environ.get("SEARXNG_URL", "http://searxng:8080").rstrip("/")
|
||||
# FLUX-Render kann bis ~90s dauern, beim ersten Render nach Container-Start
|
||||
# laedt die flux-bridge zudem ~24 GB Modell von HF (~5-10 min). Brain wartet
|
||||
# synchron — Stefan kuendigt es vorher an wenn er weiss dass es feuert.
|
||||
@@ -66,6 +70,65 @@ def _load_flux_config() -> dict:
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
def _web_search(query: str, max_results: int = 5) -> str:
|
||||
"""Fragt die self-hosted SearXNG-Instanz (JSON-API) und gibt die Top-Treffer
|
||||
als kompakten Text zurueck (Titel + Snippet + URL). Nie werfen — Fehler als
|
||||
Text-Resultat, damit der Tool-Loop weitermachen kann."""
|
||||
try:
|
||||
params = urllib.parse.urlencode({
|
||||
"q": query, "format": "json", "language": "de", "safesearch": "0",
|
||||
})
|
||||
req = urllib.request.Request(
|
||||
f"{SEARXNG_URL}/search?{params}",
|
||||
headers={"User-Agent": "ARIA/1.0", "Accept": "application/json"},
|
||||
)
|
||||
with urllib.request.urlopen(req, timeout=15) as resp:
|
||||
data = json.loads(resp.read().decode("utf-8", "ignore"))
|
||||
except Exception as exc:
|
||||
logger.warning("web_search (SearXNG) fehlgeschlagen: %s", exc)
|
||||
return f"FEHLER: Websuche nicht verfuegbar ({exc})."
|
||||
results = (data.get("results") or [])[:max_results]
|
||||
if not results:
|
||||
answers = data.get("answers") or []
|
||||
if answers:
|
||||
return "Direkte Antwort: " + " | ".join(str(a) for a in answers[:3])
|
||||
return f"Keine Web-Treffer fuer '{query}'."
|
||||
lines = [f"{len(results)} Web-Treffer fuer '{query}':"]
|
||||
for i, r in enumerate(results, 1):
|
||||
title = (r.get("title") or "").strip()
|
||||
url = (r.get("url") or "").strip()
|
||||
snippet = (r.get("content") or "").strip()
|
||||
lines.append(f"\n{i}. {title}\n {snippet[:300]}\n Quelle: {url}")
|
||||
return "\n".join(lines)
|
||||
|
||||
|
||||
# web_search ist bewusst LOCAL-ONLY (nicht in META_TOOLS): Claude hat seine
|
||||
# eigenen, staerkeren Web-Tools (WebSearch + WebFetch/Voll-Seiten-lesen + Bash).
|
||||
# SearXNG ist fuer das lokale Tier, das sonst gar keinen Netz-Zugriff haette.
|
||||
# _dispatch_tool behandelt "web_search" trotzdem generisch (Name-Match).
|
||||
WEB_SEARCH_TOOL = {
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "web_search",
|
||||
"description": (
|
||||
"Durchsuche das Web (via SearXNG) nach AKTUELLEN, nachschlagbaren "
|
||||
"Infos: Wetter, News, Fakten, Oeffnungszeiten, Preise, Definitionen. "
|
||||
"Nutze das, wenn die Antwort aktuelles Wissen braucht, das nicht im "
|
||||
"Gedaechtnis steht. Praezise Suchanfrage (Suchmaschinen-Stil). "
|
||||
"Ergebnis = Titel + Snippet + URL; fasse daraus knapp zusammen."
|
||||
),
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"query": {"type": "string", "description": "Suchanfrage (praezise)"},
|
||||
"max_results": {"type": "integer", "description": "Anzahl Treffer (Default 5, max 10)"},
|
||||
},
|
||||
"required": ["query"],
|
||||
},
|
||||
},
|
||||
}
|
||||
|
||||
|
||||
# Meta-Tool: ARIA kann selbst neue Skills bauen
|
||||
META_TOOLS = [
|
||||
{
|
||||
@@ -88,6 +151,17 @@ META_TOOLS = [
|
||||
"Stefan sich alle 60min manuell neu einloggen.\n"
|
||||
" - Bei konfigurierbaren Werten (User-IDs, Endpoints, Defaults): "
|
||||
"ueber `config_schema` deklarieren, NICHT hardcoden.\n\n"
|
||||
"SEMANTISCH BAUEN (wichtig fuer Zuverlaessigkeit!): Ein Skill soll "
|
||||
"KLARE Operationen als args anbieten, NICHT die rohe API "
|
||||
"durchreichen. Schlecht: args = {path, method, body} — dann muss das "
|
||||
"aufrufende LLM die ganze fremde API selbst kennen und baut bei "
|
||||
"komplexen Faellen (z.B. Geraete-Transfer) falsche Calls. Gut: args "
|
||||
"= {action: 'play'|'pause'|'next'|'play_on_device', device_name?: str} "
|
||||
"— der Skill-Code loest intern auf (Geraet-Name → ID, richtiger "
|
||||
"Endpoint) und kapselt die API. Faustregel: Was das LLM sonst RATEN "
|
||||
"muesste (Endpunkte, IDs, Payload-Struktur), gehoert INS Skill. Das "
|
||||
"`args`-Schema ist die Bedienungsanleitung, die das LLM sieht — mach "
|
||||
"sie semantisch und selbsterklaerend.\n\n"
|
||||
"HARTE REGEL — IMMER Skill anlegen wenn: die Loesung erfordert eine "
|
||||
"pip-Library. Sonst muesste der Install bei jedem Container-Restart "
|
||||
"neu laufen (Brain hat keinen persistenten State ausser /data/skills/).\n\n"
|
||||
@@ -136,17 +210,28 @@ META_TOOLS = [
|
||||
"description": (
|
||||
"OPTIONAL — fuer 'reines Steuern'-Skills (Licht an/aus, Spotify "
|
||||
"pause/next, Rollade hoch/runter etc.) eine Liste von "
|
||||
"[{match, args, reply}] eintragen. Wenn ein User-Befehl gegen "
|
||||
"match (anchored Regex, case-insensitive) matched, ruft das "
|
||||
"Brain run_skill(name, args) DIREKT auf und gibt reply zurueck — "
|
||||
"ohne Claude (~5s Latenz gespart). Match wird gegen den "
|
||||
"normalisierten Text (lowercase, Endsatzzeichen weg) gemacht; "
|
||||
"schreibe Patterns mit ^...$ damit nur exakte Befehle matchen "
|
||||
"und nicht Teilstrings (z.B. ^pause$ statt pause). NICHT fuer "
|
||||
"Skills mit kreativem Output / parametrisierter Logik — die "
|
||||
"brauchen Claude. Beispiel: "
|
||||
"[{\"match\":\"^pause$\",\"args\":{\"path\":\"/v1/me/player/pause\",\"method\":\"PUT\"},"
|
||||
"\"reply\":\"Spotify: pausiert ⏸\"}]"
|
||||
"[{match, args, reply}]. Bei match (anchored Regex, "
|
||||
"case-insensitive, gegen den normalisierten Text: lowercase, "
|
||||
"Endsatzzeichen weg) ruft das Brain run_skill(name, args) DIREKT "
|
||||
"auf und gibt reply zurueck — ohne LLM (instant).\n\n"
|
||||
"ROBUST SCHREIBEN (wichtig!):\n"
|
||||
"- Immer ^...$ (nur ganze Befehle, keine Teilstrings).\n"
|
||||
"- **Wortstellung + Synonyme + Fuellwoerter abdecken** in EINEM "
|
||||
"Pattern via Alternativen und optionalen Gruppen. Nicht nur die "
|
||||
"eine Formulierung! Beispiel Pause: "
|
||||
"`^(spotify |musik )?(pause|pausier(e|en)?|stop|stopp|halt|"
|
||||
"anhalten)( mal| bitte| spotify| die musik)?$` faengt 'pause', "
|
||||
"'pause spotify', 'stopp mal', 'musik anhalten' … alle ab.\n"
|
||||
"- Optionale Fuellwoerter mit `( bitte| mal| doch)?` zulassen, "
|
||||
"Artikel/Objekte mit `( das lied| den song| die musik)?`.\n\n"
|
||||
"Aber KEIN Zwang zur Vollstaendigkeit: Fast-Paths sind nur fuer "
|
||||
"die HAEUFIGSTEN exakten Befehle (instant). Seltene/ungewoehnliche "
|
||||
"Formulierungen faengt das lokale LLM ohnehin schnell ab (es hat "
|
||||
"das Tool) — also lieber ein paar solide, breite Patterns als 30 "
|
||||
"enge. NICHT fuer Skills mit kreativem/parametrisiertem Output. "
|
||||
"Beispiel: [{\"match\":\"^(spotify |musik )?(pause|stop|stopp)"
|
||||
"( mal| bitte)?$\",\"args\":{\"path\":\"/v1/me/player/pause\","
|
||||
"\"method\":\"PUT\"},\"reply\":\"Spotify: pausiert ⏸\"}]"
|
||||
),
|
||||
},
|
||||
},
|
||||
@@ -933,10 +1018,36 @@ META_TOOLS = [
|
||||
# Diese Logik ist generisch — ARIA deklariert die Patterns selbst beim
|
||||
# skill_create / skill_update, das Brain orchestriert nur.
|
||||
|
||||
def _strip_leading_hint_blocks(text: str) -> str:
|
||||
"""Entfernt fuehrende Hint-Bloecke `[ ... ]`, die die Bridge an
|
||||
Voice-Nachrichten haengt (GPS-Position, Barge-In-Hinweis). Ohne das matcht
|
||||
KEIN Voice-Befehl je den Fast-Path (Regex ist ^...$-verankert) — selbst
|
||||
'nächstes lied' landete unnoetig bei Claude statt beim 0-Token-Fast-Path."""
|
||||
s = (text or "").strip()
|
||||
prev = None
|
||||
while prev != s:
|
||||
prev = s
|
||||
s = re.sub(r"^\s*\[[^\]]*\]\s*", "", s)
|
||||
return s
|
||||
|
||||
|
||||
def _fold_umlauts(s: str) -> str:
|
||||
"""ä→ae, ö→oe, ü→ue, ß→ss (lowercase erwartet). Whisper liefert echte
|
||||
Umlaute ('Nächstes Lied'), viele Skill-fast_patterns sind aber in ae/oe/ue
|
||||
geschrieben — ohne Faltung matcht das nie."""
|
||||
return (s.replace("ä", "ae").replace("ö", "oe").replace("ü", "ue")
|
||||
.replace("ß", "ss"))
|
||||
|
||||
|
||||
def _normalize_for_fast_match(text: str) -> str:
|
||||
norm = (text or "").strip().lower()
|
||||
norm = re.sub(r"[.!?]+$", "", norm)
|
||||
norm = re.sub(r"\s+", " ", norm)
|
||||
norm = _strip_leading_hint_blocks(text).lower()
|
||||
norm = _fold_umlauts(norm)
|
||||
# Interne Satzzeichen raus, die Whisper einstreut ('Nächstes Lied, bitte.').
|
||||
# Kommas/Semikola/Doppelpunkte → Space, Punkt/!/? ganz weg. Sonst matcht das
|
||||
# ^...$-verankerte fast_pattern nie, weil das Komma dazwischenfunkt.
|
||||
norm = re.sub(r"[,;:]", " ", norm)
|
||||
norm = re.sub(r"[.!?]+", "", norm)
|
||||
norm = re.sub(r"\s+", " ", norm).strip()
|
||||
return norm
|
||||
|
||||
|
||||
@@ -1032,6 +1143,9 @@ class Agent:
|
||||
rx = pat.get("match") or ""
|
||||
if not rx:
|
||||
continue
|
||||
# Pattern GLEICH falten wie den Text — egal ob der Skill-Autor
|
||||
# 'naechstes' oder 'nächstes' geschrieben hat.
|
||||
rx = _fold_umlauts(rx)
|
||||
try:
|
||||
if not re.match(rx, norm, re.IGNORECASE):
|
||||
continue
|
||||
@@ -1055,60 +1169,158 @@ class Agent:
|
||||
return reply
|
||||
return None
|
||||
|
||||
# ── Fast-Lane: lokales schnelles LLM (Plan B, B1a) ──
|
||||
# ── Fast-Lane: lokales schnelles LLM (Plan B, B1a/B1b) ──
|
||||
#
|
||||
# Zwischen Skill-Fast-Path und Claude-Loop: einfache Plauder-Turns beantwortet
|
||||
# das lokale Qwen (schlanker Prompt, KEINE Tools) in <1 s. Gated ueber
|
||||
# Zwischen Skill-Fast-Path und Claude-Loop: einfache Turns beantwortet das
|
||||
# lokale Qwen in <1 s. Seit B1b mit kuratierten Tools (web_search,
|
||||
# memory_search, trigger_timer, Spotify). Gated ueber
|
||||
# /shared/config/local_llm.json (Default aus → alles Claude wie bisher).
|
||||
# Rueckgabe: fertige Antwort (str) wenn lokal erledigt, sonst None → Claude.
|
||||
|
||||
_LOCAL_WINDOW_TURNS = 8 # nur die letzten N Turns ans lokale Modell (Speed)
|
||||
_LOCAL_WINDOW_TURNS = 8 # nur die letzten N Turns ans lokale Modell (Speed)
|
||||
_LOCAL_TOOL_ITERATIONS = 3 # max Tool-Runden lokal, sonst → Claude
|
||||
|
||||
# Weiche Fehler-Marker im Tool-Ausgabetext. Viele Action-Skills geben einen
|
||||
# menschenlesbaren Fehler auf stdout aus und beenden sich TROTZDEM mit Exit 0
|
||||
# (z.B. Spotify: "Fehler beim Uebertragen", "Geraet nicht gefunden"). Der
|
||||
# harte FEHLER-Prefix (Exit != 0) faengt das nicht. Fuer run_*-Skills werten
|
||||
# wir daher auch solche Marker als Fehlschlag → eskalieren an Claude, der
|
||||
# mehrstufig weiterdenkt. Info-Tools (web_search/memory_search) sind bewusst
|
||||
# AUSGENOMMEN: deren Inhalt darf das Wort "Fehler" tragen ohne dass der
|
||||
# Tool-Call scheiterte. Eskalieren ist immer sicher (nur langsamer).
|
||||
_SOFT_FAIL_MARKERS = (
|
||||
"fehler", "fehlgeschlagen", "nicht gefunden", "not found",
|
||||
"konnte nicht", "keine verbindung", "nicht verfuegbar",
|
||||
"exception", "traceback",
|
||||
)
|
||||
|
||||
def _local_tool_failed(self, tool_name: str, tresult: str) -> bool:
|
||||
"""True wenn ein lokaler Tool-Call als gescheitert gilt (→ Claude)."""
|
||||
s = (tresult or "").strip()
|
||||
if not s:
|
||||
return False
|
||||
if s.startswith("FEHLER"): # harter Exit-Code-Fehler
|
||||
return True
|
||||
if tool_name.startswith("run_"): # Action-Skill: auch weiche Fehler
|
||||
low = s.lower()
|
||||
return any(mk in low for mk in self._SOFT_FAIL_MARKERS)
|
||||
return False
|
||||
|
||||
# Kuratierte Tool-Auswahl fuers lokale Tier (B1b): web_search (local-only,
|
||||
# SearXNG) + memory_search/trigger_timer (aus META_TOOLS) + Spotify-Skill.
|
||||
# Bewusst klein (Speed + Sicherheit); alles andere → Claude.
|
||||
_LOCAL_TOOL_NAMES = {"memory_search", "trigger_timer"}
|
||||
|
||||
def _build_local_tools(self) -> list:
|
||||
tools = [WEB_SEARCH_TOOL]
|
||||
tools += [t for t in META_TOOLS
|
||||
if t.get("function", {}).get("name") in self._LOCAL_TOOL_NAMES]
|
||||
for s in skills_mod.list_skills(active_only=False):
|
||||
if s.get("name") == "spotify" and s.get("active", True):
|
||||
tools.append(_skill_to_tool(s))
|
||||
break
|
||||
return tools
|
||||
|
||||
def _try_local_fast_lane(self, user_message: str,
|
||||
active_project_id: str) -> Optional[str]:
|
||||
cfg = router_mod.load_config()
|
||||
if not router_mod.should_try_local(user_message, cfg):
|
||||
return None
|
||||
local_only = bool(cfg.get("localOnly"))
|
||||
local_model = cfg.get("localLlmModel") or "qwen3-8b" # B0.5: llama-swap-Key
|
||||
tools = self._build_local_tools() # B1b: kuratierte Tools
|
||||
|
||||
sys_prompt = router_mod.build_local_system_prompt(IDENTITY_ANCHOR)
|
||||
# Nur die letzten paar Turns ans lokale Modell — es ist fuer kurze
|
||||
# Plauder-Turns da. Volles Fenster (bis 50) wuerde das Prefill aufblaehen
|
||||
# und den Speed-Vorteil auffressen (gemessen: 12 Turns → ~2,6s statt ~0,8s).
|
||||
sys_prompt = router_mod.build_local_system_prompt(IDENTITY_ANCHOR,
|
||||
has_tools=bool(tools))
|
||||
# Nur die letzten paar Turns ans lokale Modell (Speed — volles Fenster
|
||||
# wuerde das Prefill aufblaehen).
|
||||
window = self.conversation.window(project_id=active_project_id)[-self._LOCAL_WINDOW_TURNS:]
|
||||
messages = [{"role": "system", "content": sys_prompt}]
|
||||
messages += [{"role": t.role, "content": t.content} for t in window]
|
||||
|
||||
res = local_llm_chat(messages, max_tokens=400, temperature=0.5)
|
||||
local_only = bool(cfg.get("localOnly"))
|
||||
# Tool-Loop: lokales Modell darf web_search/memory_search/trigger_timer/
|
||||
# Spotify aufrufen. Ergebnisse zurueck, bis es final (ohne tool_calls) antwortet.
|
||||
final = ""
|
||||
for _ in range(self._LOCAL_TOOL_ITERATIONS):
|
||||
res = local_llm_chat(messages, max_tokens=500, temperature=0.5,
|
||||
tools=tools, model=local_model)
|
||||
if not res.get("ok"):
|
||||
logger.info("[router] lokal fehlgeschlagen (%s) — %s", res.get("error"),
|
||||
"kein Fallback (localOnly)" if local_only else "→ Claude")
|
||||
if local_only:
|
||||
return f"[Lokales LLM nicht erreichbar: {res.get('error', 'unbekannt')}]"
|
||||
return None
|
||||
# Metric: dieser lokale Call (echte usage-Tokens wenn der Adapter sie
|
||||
# liefert). Erfasst pro Tool-Runde — mehrere Runden = mehrere Calls.
|
||||
try:
|
||||
metrics.log_local_call(res.get("model") or local_model, messages,
|
||||
res.get("content") or "", res.get("usage"))
|
||||
except Exception:
|
||||
pass
|
||||
tcs = res.get("tool_calls")
|
||||
if tcs:
|
||||
messages.append({"role": "assistant",
|
||||
"content": res.get("content") or "",
|
||||
"tool_calls": tcs})
|
||||
had_error = False
|
||||
for tc in tcs:
|
||||
fn = tc.get("function") or {}
|
||||
tname = fn.get("name") or ""
|
||||
try:
|
||||
targs = json.loads(fn.get("arguments") or "{}")
|
||||
except Exception:
|
||||
targs = {}
|
||||
logger.info("[router] lokal Tool-Call: %s(%s)", tname,
|
||||
", ".join(targs.keys()))
|
||||
tresult = self._dispatch_tool(tname, targs)
|
||||
if self._local_tool_failed(tname, tresult):
|
||||
had_error = True
|
||||
messages.append({"role": "tool",
|
||||
"tool_call_id": tc.get("id") or "",
|
||||
"name": tname,
|
||||
"content": (tresult or "")[:6000]})
|
||||
# GENERAL (kein per-Skill-Code): scheitert ein Tool-Call, macht
|
||||
# das grosse Modell weiter — es baut komplexe/rohe API-Calls
|
||||
# zuverlaessiger und behandelt Fehler besser. So muss der Router
|
||||
# NICHT wissen, welche Skill-Aufrufe "schwer" sind; das lokale
|
||||
# Tier probiert, und bei Fehler uebernimmt Claude.
|
||||
if had_error and not local_only:
|
||||
logger.info("[router] lokaler Tool-Fehler → Claude uebernimmt")
|
||||
return None
|
||||
continue # naechste Runde mit Tool-Ergebnissen
|
||||
final = (res.get("content") or "").strip()
|
||||
break
|
||||
else:
|
||||
logger.info("[router] lokal Tool-Loop-Limit → %s",
|
||||
"Fallback (localOnly)" if local_only else "Claude")
|
||||
if not local_only:
|
||||
return None
|
||||
final = final or "[Lokales LLM: Tool-Loop-Limit erreicht.]"
|
||||
|
||||
if not res.get("ok"):
|
||||
logger.info("[router] lokal fehlgeschlagen (%s) — %s",
|
||||
res.get("error"), "kein Fallback (localOnly)" if local_only else "→ Claude")
|
||||
if local_only:
|
||||
# Eval-Modus: KEIN Claude. Ehrliche Fehlermeldung statt Stille.
|
||||
return f"[Lokales LLM nicht erreichbar: {res.get('error', 'unbekannt')}]"
|
||||
return None
|
||||
|
||||
content = (res.get("content") or "").strip()
|
||||
# Sicherheitsnetz: das lokale Modell wickelt seine Antwort manchmal
|
||||
# faelschlich komplett in <voice>...</voice> (aus dem Kontext imitiert).
|
||||
# Das wuerde die Anzeige leeren (Display strippt <voice>). Tags raus —
|
||||
# der lokale Reply ist kurz, Plain-Text dient Anzeige UND TTS.
|
||||
final = re.sub(r"</?voice>", "", final).strip()
|
||||
|
||||
if local_only:
|
||||
# Erzwungen lokal: Escalation-Marker ignorieren, Antwort so nehmen.
|
||||
content = content.replace(router_mod.ESCALATE_MARKER, "").strip()
|
||||
if not content:
|
||||
final = final.replace(router_mod.ESCALATE_MARKER, "").strip()
|
||||
if not final:
|
||||
return "[Lokales LLM lieferte keine Antwort.]"
|
||||
logger.info("[router] lokal (localOnly) %sms", res.get("elapsedMs"))
|
||||
self.conversation.add("assistant", content, project_id=active_project_id)
|
||||
return content
|
||||
logger.info("[router] lokal (localOnly) beantwortet")
|
||||
self.conversation.add("assistant", final, project_id=active_project_id)
|
||||
return final
|
||||
|
||||
# Normalbetrieb: leere Antwort oder Escalation-Marker → Claude.
|
||||
if not content or router_mod.ESCALATE_MARKER in content:
|
||||
# Normalbetrieb: leer, Escalation-Marker oder (unwahrscheinlich) ein
|
||||
# Identity-Break → Claude uebernehmen.
|
||||
if (not final or router_mod.ESCALATE_MARKER in final
|
||||
or looks_like_identity_break(final)):
|
||||
logger.info("[router] lokal eskaliert → Claude")
|
||||
return None
|
||||
|
||||
logger.info("[router] lokal beantwortet in %sms (%d Zeichen)",
|
||||
res.get("elapsedMs"), len(content))
|
||||
self.conversation.add("assistant", content, project_id=active_project_id)
|
||||
return content
|
||||
logger.info("[router] lokal beantwortet (%d Zeichen)", len(final))
|
||||
self.conversation.add("assistant", final, project_id=active_project_id)
|
||||
return final
|
||||
|
||||
# ── Hauptpfad: ein User-Turn → Tool-Loop → finaler Reply ──
|
||||
|
||||
@@ -1116,7 +1328,7 @@ class Agent:
|
||||
|
||||
def chat(self, user_message: str, source: str = "",
|
||||
project_id: Optional[str] = None,
|
||||
pending_queue: Optional[list[str]] = None) -> str:
|
||||
pending_queue: Optional[list[str]] = None) -> tuple:
|
||||
"""Verarbeitet eine User-Nachricht — pro Request project_id explizit
|
||||
angegeben (leer = Hauptchat). Kein globaler active_project-State mehr —
|
||||
so laufen parallele /chat-Requests fuer verschiedene Projekte echt
|
||||
@@ -1150,7 +1362,14 @@ class Agent:
|
||||
self.conversation.add("assistant", fast_reply, project_id=active_project_id)
|
||||
if active_project_id:
|
||||
projects_mod.touch_project(active_project_id)
|
||||
return fast_reply
|
||||
# Metric: Fast-Path spart einen ganzen Claude-Call zum Nulltarif.
|
||||
try:
|
||||
metrics.log_fast_path(fast_reply)
|
||||
except Exception:
|
||||
pass
|
||||
# Fast-Path = reiner Steuerbefehl → NICHT vorlesen (speak=False).
|
||||
# System-Flag statt <voice>-Tag: robust, unabhaengig vom Skill-Inhalt.
|
||||
return fast_reply, "fast-path", False
|
||||
|
||||
# 1. User-Turn an die Konversation
|
||||
self.conversation.add("user", user_message, source=source,
|
||||
@@ -1164,7 +1383,7 @@ class Agent:
|
||||
# teure Claude-Aufbau + Tool-Loop wird uebersprungen. Sonst None → Claude.
|
||||
local_reply = self._try_local_fast_lane(user_message, active_project_id)
|
||||
if local_reply is not None:
|
||||
return local_reply
|
||||
return local_reply, "local", True # ARIA-Antwort → vorlesen ok
|
||||
|
||||
# 2. Hot Memory (alle pinned Punkte)
|
||||
hot = self.store.list_pinned()
|
||||
@@ -1339,10 +1558,33 @@ class Agent:
|
||||
logger.warning("Konnte Error-Marker nicht persistieren: %s", add_exc)
|
||||
raise
|
||||
|
||||
# Gift-Waechter: faellt Claude trotz --system-prompt + Seed aus der Rolle
|
||||
# (Identity-Break), NICHT persistieren — sonst vergiftet dieser eine Turn
|
||||
# die History und loest bei schwachen Folgeturns eine Kaskade aus (das
|
||||
# Modell setzt seine eigene Ablehnung fort). Ein Retry holt per
|
||||
# Nondeterminismus meist die ARIA-Antwort; sonst sichere Fallback-Antwort.
|
||||
# So kann ein einzelner Ausrutscher nie snowballen.
|
||||
if looks_like_identity_break(final_reply):
|
||||
logger.warning("[guard] Identity-Break in Antwort erkannt — Retry")
|
||||
try:
|
||||
retry = self.proxy.chat_full(messages, tools=tools,
|
||||
project_id=active_project_id)
|
||||
retry_text = (retry.content or "").strip()
|
||||
except Exception as exc:
|
||||
logger.warning("[guard] Retry fehlgeschlagen: %s", exc)
|
||||
retry_text = ""
|
||||
if retry_text and not looks_like_identity_break(retry_text):
|
||||
logger.info("[guard] Retry lieferte saubere Antwort")
|
||||
final_reply = retry_text
|
||||
else:
|
||||
logger.warning("[guard] Retry weiter Break/leer — Fallback, Break NICHT persistiert")
|
||||
final_reply = ("Hey, ich bin ARIA. \U0001F60A Bei mir ist alles bereit — "
|
||||
"sag mir einfach, was du brauchst.")
|
||||
|
||||
# 7. Assistant-Turn (final reply) in die Conversation
|
||||
self.conversation.add("assistant", final_reply,
|
||||
project_id=active_project_id)
|
||||
return final_reply
|
||||
return final_reply, "claude", True # ARIA-Antwort → vorlesen ok
|
||||
|
||||
# ── Tool-Dispatcher ───────────────────────────────────────
|
||||
|
||||
@@ -1783,6 +2025,15 @@ class Agent:
|
||||
except Exception as e:
|
||||
logger.exception("memory_search fehlgeschlagen")
|
||||
return f"FEHLER: {e}"
|
||||
if name == "web_search":
|
||||
query = (arguments.get("query") or "").strip()
|
||||
if not query:
|
||||
return "FEHLER: query ist Pflicht."
|
||||
try:
|
||||
n = int(arguments.get("max_results", 5))
|
||||
except (TypeError, ValueError):
|
||||
n = 5
|
||||
return _web_search(query, max(1, min(n, 10)))
|
||||
if name == "memory_update":
|
||||
pid = (arguments.get("id") or "").strip()
|
||||
if not pid:
|
||||
|
||||
@@ -150,7 +150,7 @@ async def _fire(trigger: dict, agent_factory) -> None:
|
||||
|
||||
try:
|
||||
agent = agent_factory()
|
||||
reply = agent.chat(prompt, source="trigger")
|
||||
reply, _, _ = agent.chat(prompt, source="trigger")
|
||||
events = agent.pop_events()
|
||||
logger.info("[trigger] %s gefeuert → ARIA-Reply: %s", name, reply[:80])
|
||||
triggers_mod.append_log(name, {"event": "reply", "text": reply[:500]})
|
||||
|
||||
@@ -41,19 +41,28 @@ import shutil
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
# "claude code" ist fuer sich genommen noch kein Beweis (Stefan und ARIA reden
|
||||
# im Dev-Kontext legitim ueber Claude Code). Erst in Kombination mit einem
|
||||
# zweiten Ablehnungs-Marker ist es eindeutig eine aus-der-Rolle-Antwort.
|
||||
_PRIMARY = re.compile(r"claude\s*code", re.IGNORECASE)
|
||||
_SECONDARY = re.compile(
|
||||
r"injiz|inject|fabriz|fabricat|adoptier|adopting|"
|
||||
r"prompt[\s-]*injection|keine echten|nicht (?:real|adopt)",
|
||||
# STARKE, selbstreferenzielle Break-Marker — identisch zu prompts._IDENTITY_BREAK
|
||||
# (dem Laufzeit-Gift-Waechter). Hier dupliziert, damit das Script self-contained
|
||||
# ist (laeuft auch auf dem Host-Python ohne qdrant/prompts-Import). Bewusst NICHT
|
||||
# das blosse Wort "injizier"/"prompt injection" — das nutzt ARIA in Pentest-
|
||||
# Antworten legitim (sonst False Positives auf echte Security-Doku, wie im
|
||||
# Dry-Run gesehen: "Runde 60 … SSRF", "Dein Ziel: LLM …").
|
||||
_BREAK = re.compile(
|
||||
r"ich\s+bin\s+(?:allerdings\s+|ja\s+|nach\s+wie\s+vor\s+|weiterhin\s+)*claude|"
|
||||
r"i'?m\s+(?:still\s+|actually\s+)?claude\s+code|i\s+am\s+claude\b|"
|
||||
r"erfundene[nr]?\s+(?:tool|persona|schemas)|fabricated\s+persona|"
|
||||
r"fabrizierte?\s+(?:persona|gespr|konversation)|fabricated\s+conversation|"
|
||||
r"fake[- ]persona|injizierte[rn]?\s+(?:system-?prompt|kontext|persona)|"
|
||||
r"injected\s+(?:system\s*prompt|persona|context)|"
|
||||
r"diese\s+session\s+enthält\s+(?:einen|eine)\b.{0,40}injizier|"
|
||||
r"this\s+session\s+(?:contains|has|keeps|repeatedly)\b.{0,40}(?:inject|fabricat|fake)|"
|
||||
r"nicht\s+real\s+in\s+dieser\s+(?:umgebung|session)|not\s+real\s+in\s+this",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
|
||||
|
||||
def is_poison(content: str) -> bool:
|
||||
return bool(_PRIMARY.search(content) and _SECONDARY.search(content))
|
||||
return bool(_BREAK.search(content or ""))
|
||||
|
||||
|
||||
def get_content(obj: dict) -> str:
|
||||
|
||||
@@ -29,14 +29,21 @@ LOCAL_LLM_HTTP_TIMEOUT_SEC = float(os.environ.get("LOCAL_LLM_HTTP_TIMEOUT_SEC",
|
||||
|
||||
|
||||
def local_llm_chat(messages: list, *, max_tokens: int = 512,
|
||||
temperature: float = 0.7, stop=None) -> dict:
|
||||
temperature: float = 0.7, stop=None, tools=None,
|
||||
model=None) -> dict:
|
||||
"""Ein Chat-Call ans lokale LLM. messages = [{role, content}, ...].
|
||||
Blockierend (urllib) — im Brain laeuft chat() ohnehin im Executor-Thread."""
|
||||
model (B0.5): welches Modell llama-swap laden soll. tools (B1b): optionale
|
||||
OpenAI-Tool-Defs; das Ergebnis kann dann result['tool_calls'] enthalten.
|
||||
Blockierend (urllib) — chat() laeuft ohnehin im Executor-Thread."""
|
||||
if not isinstance(messages, list) or not messages:
|
||||
return {"ok": False, "error": "messages leer/ungueltig"}
|
||||
req = {"messages": messages, "max_tokens": max_tokens, "temperature": temperature}
|
||||
if stop:
|
||||
req["stop"] = stop
|
||||
if tools:
|
||||
req["tools"] = tools
|
||||
if model:
|
||||
req["model"] = model
|
||||
try:
|
||||
body = json.dumps(req).encode("utf-8")
|
||||
http_req = urllib.request.Request(
|
||||
|
||||
+10
-1
@@ -630,6 +630,12 @@ class ChatOut(BaseModel):
|
||||
turns: int
|
||||
distilling: bool
|
||||
events: list = Field(default_factory=list)
|
||||
# Welcher Backend die Antwort erzeugt hat: "local" (Qwen), "claude",
|
||||
# "fast-path" (Skill/Regex). Fuer den Quell-Badge in Diagnostic.
|
||||
answered_by: str = "claude"
|
||||
# Soll die Antwort vorgelesen werden? Fast-Path (reiner Steuerbefehl) = False;
|
||||
# ARIA-Antworten (local/claude) = True. System-Flag statt <voice>-Tag.
|
||||
speak: bool = True
|
||||
# Echo der project_id die dieser Turn hatte. Bridge nutzt sie damit die
|
||||
# ausgehende Chat-Bubble sauber getaggt in der richtigen Thread-Bahn der
|
||||
# UI landet.
|
||||
@@ -713,7 +719,7 @@ async def chat(body: ChatIn, background: BackgroundTasks):
|
||||
# Sync-Aufruf im Executor damit wir den Event-Loop nicht blocken —
|
||||
# chat() macht HTTP-Calls (Proxy) die 30-60s dauern koennen.
|
||||
loop = asyncio.get_running_loop()
|
||||
reply = await loop.run_in_executor(
|
||||
reply, answered_by, speak = await loop.run_in_executor(
|
||||
None,
|
||||
lambda: a.chat(
|
||||
body.message, source=body.source, project_id=pid,
|
||||
@@ -735,6 +741,8 @@ async def chat(body: ChatIn, background: BackgroundTasks):
|
||||
distilling=needs_distill,
|
||||
events=a.pop_events(),
|
||||
project_id=pid,
|
||||
answered_by=answered_by,
|
||||
speak=speak,
|
||||
)
|
||||
finally:
|
||||
_project_pending[pid] = [
|
||||
@@ -809,6 +817,7 @@ def projects_archive(project_id: str):
|
||||
class ProjectUpdateBody(BaseModel):
|
||||
name: Optional[str] = None
|
||||
description: Optional[str] = None
|
||||
hidden: Optional[bool] = None
|
||||
|
||||
|
||||
@app.patch("/projects/{project_id}")
|
||||
|
||||
+54
-10
@@ -52,25 +52,55 @@ def _messages_tokens(messages: list) -> int:
|
||||
return total
|
||||
|
||||
|
||||
def log_call(model: str, messages_in: list, reply_text: str = "") -> None:
|
||||
"""Eine Call-Metric anhaengen. Robust gegen Fehler (silent fail)."""
|
||||
def _append(model: str, tokens_in: int, tokens_out: int, source: str) -> None:
|
||||
"""Ein Metric-Entry auf Disk anhaengen. Robust (silent fail)."""
|
||||
try:
|
||||
tokens_in = _messages_tokens(messages_in)
|
||||
tokens_out = _estimate_tokens(reply_text)
|
||||
line = json.dumps({
|
||||
"ts": int(time.time() * 1000),
|
||||
"model": model,
|
||||
"in": tokens_in,
|
||||
"out": tokens_out,
|
||||
"in": int(tokens_in),
|
||||
"out": int(tokens_out),
|
||||
"source": source, # "claude" | "local" | "fast-path"
|
||||
})
|
||||
METRICS_FILE.parent.mkdir(parents=True, exist_ok=True)
|
||||
with METRICS_FILE.open("a", encoding="utf-8") as f:
|
||||
f.write(line + "\n")
|
||||
# Sanftes Rotate ohne hohe IO-Kosten — nur alle 1000 Calls checken
|
||||
if (tokens_in + tokens_out) % 1000 < 4:
|
||||
_maybe_rotate()
|
||||
except Exception as exc:
|
||||
logger.warning("metrics.log_call: %s", exc)
|
||||
logger.warning("metrics._append: %s", exc)
|
||||
|
||||
|
||||
def log_call(model: str, messages_in: list, reply_text: str = "",
|
||||
source: str = "claude") -> None:
|
||||
"""Claude-Call-Metric anhaengen (Tokens per chars/4-Schaetzung)."""
|
||||
_append(model, _messages_tokens(messages_in), _estimate_tokens(reply_text), source)
|
||||
|
||||
|
||||
def log_local_call(model: str, messages_in: list, reply_text: str = "",
|
||||
usage: dict | None = None) -> None:
|
||||
"""Lokaler-LLM-Call-Metric. Nutzt echte usage-Tokens (prompt/completion)
|
||||
wenn der Adapter sie liefert, sonst chars/4-Schaetzung wie bei Claude.
|
||||
Quelle = 'local' — damit die Ersparnis-Rechnung local von claude trennt."""
|
||||
tokens_in = tokens_out = None
|
||||
if isinstance(usage, dict):
|
||||
pt = usage.get("prompt_tokens")
|
||||
ct = usage.get("completion_tokens")
|
||||
if isinstance(pt, (int, float)):
|
||||
tokens_in = int(pt)
|
||||
if isinstance(ct, (int, float)):
|
||||
tokens_out = int(ct)
|
||||
if tokens_in is None:
|
||||
tokens_in = _messages_tokens(messages_in)
|
||||
if tokens_out is None:
|
||||
tokens_out = _estimate_tokens(reply_text)
|
||||
_append(model or "local", tokens_in, tokens_out, "local")
|
||||
|
||||
|
||||
def log_fast_path(reply_text: str = "") -> None:
|
||||
"""Fast-Path (reiner Skill, KEIN LLM) — spart einen ganzen Claude-Call zum
|
||||
Nulltarif. tokens_in=0 (kein Prompt ans LLM), out = winzige Quittung."""
|
||||
_append("fast-path", 0, _estimate_tokens(reply_text), "fast-path")
|
||||
|
||||
|
||||
def _maybe_rotate() -> None:
|
||||
@@ -95,6 +125,11 @@ def aggregate(window_seconds: int) -> dict:
|
||||
tokens_in = 0
|
||||
tokens_out = 0
|
||||
by_model: dict[str, int] = {}
|
||||
# Aufschluesselung nach Quelle (claude / local / fast-path) fuer die
|
||||
# Ersparnis-Anzeige im Diagnostic.
|
||||
def _src_bucket() -> dict:
|
||||
return {"calls": 0, "tokens_in": 0, "tokens_out": 0}
|
||||
by_source: dict[str, dict] = {}
|
||||
if METRICS_FILE.exists():
|
||||
try:
|
||||
for raw in METRICS_FILE.read_text(encoding="utf-8").splitlines():
|
||||
@@ -107,11 +142,19 @@ def aggregate(window_seconds: int) -> dict:
|
||||
continue
|
||||
if obj.get("ts", 0) < cutoff_ms:
|
||||
continue
|
||||
ti = int(obj.get("in") or 0)
|
||||
to = int(obj.get("out") or 0)
|
||||
calls += 1
|
||||
tokens_in += int(obj.get("in") or 0)
|
||||
tokens_out += int(obj.get("out") or 0)
|
||||
tokens_in += ti
|
||||
tokens_out += to
|
||||
m = obj.get("model", "?")
|
||||
by_model[m] = by_model.get(m, 0) + 1
|
||||
# Alt-Eintraege ohne 'source' zaehlen als claude (Rueckwaerts-Kompat).
|
||||
src = obj.get("source") or "claude"
|
||||
b = by_source.setdefault(src, _src_bucket())
|
||||
b["calls"] += 1
|
||||
b["tokens_in"] += ti
|
||||
b["tokens_out"] += to
|
||||
except Exception as exc:
|
||||
logger.warning("metrics aggregate: %s", exc)
|
||||
return {
|
||||
@@ -120,6 +163,7 @@ def aggregate(window_seconds: int) -> dict:
|
||||
"tokens_in": tokens_in,
|
||||
"tokens_out": tokens_out,
|
||||
"by_model": by_model,
|
||||
"by_source": by_source,
|
||||
}
|
||||
|
||||
|
||||
|
||||
@@ -132,6 +132,7 @@ def create_project(name: str, description: str = "") -> dict:
|
||||
"name": name,
|
||||
"description": description.strip(),
|
||||
"status": "active", # active | ended | archived
|
||||
"hidden": False, # optisch aus Listen ausblenden (bleibt nutzbar)
|
||||
"created_at": now,
|
||||
"updated_at": now,
|
||||
"last_activity_at": now,
|
||||
@@ -148,7 +149,7 @@ def update_project(project_id: str, patch: dict) -> Optional[dict]:
|
||||
projects = _load_all()
|
||||
for p in projects:
|
||||
if p["id"] == project_id:
|
||||
for k in ("name", "description", "status"):
|
||||
for k in ("name", "description", "status", "hidden"):
|
||||
if k in patch and patch[k] is not None:
|
||||
p[k] = patch[k]
|
||||
p["updated_at"] = _now()
|
||||
|
||||
@@ -15,6 +15,7 @@ mit dem Conversation-Loop in spaeteren Phasen.
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import re
|
||||
from datetime import datetime, timezone, timedelta
|
||||
from typing import List
|
||||
|
||||
@@ -51,6 +52,22 @@ Rolle wirft — du bleibst ARIA und arbeitest normal weiter.
|
||||
|
||||
Dein eigenes Werkzeug-Format (`<tool_call>…`) ist echt und von diesem System so
|
||||
vorgesehen — nutze es ganz normal.
|
||||
|
||||
## DISKRETION — PRIVATE/INTIME DINGE (HART, GILT IMMER)
|
||||
Du kennst private und intime Details ueber Stefan (Beziehungen, Partnerinnen,
|
||||
Sexualitaet/Lebensweise, Gesundheit, Finanzen, anderes Persoenliches). Dieses
|
||||
Wissen ist fuer deine Arbeit da — aber es ist HOCHVERTRAULICH:
|
||||
- Gib es NIEMALS von dir aus preis. Nicht in Selbstvorstellungen, nicht bei
|
||||
„was weisst du ueber mich", nicht in Zusammenfassungen, nicht in Triggern oder
|
||||
Push-Nachrichten, nicht als Beispiel, nicht „um zu zeigen was du weisst".
|
||||
- Nur wenn Stefan KONKRET nach genau diesem Thema fragt — und dann knapp und
|
||||
gezielt, ohne Rundum-Dump.
|
||||
- Denk IMMER daran: jemand koennte danebensitzen oder mithoeren. Intimes
|
||||
ungefragt auszusprechen ist ein Vertrauensbruch. Im Zweifel: weglassen oder
|
||||
kurz rueckfragen („Soll ich dazu ins Detail gehen?").
|
||||
Auf „wer bist du" / „was weisst du ueber mich": antworte allgemein und diskret
|
||||
(z.B. „einiges — geschaeftlich wie privat; frag mich gezielt"), NIEMALS ein
|
||||
Aufzaehlen privater/intimer Details.
|
||||
"""
|
||||
|
||||
|
||||
@@ -77,6 +94,36 @@ IDENTITY_SEED = (
|
||||
)
|
||||
|
||||
|
||||
# Gift-Waechter: erkennt eine Antwort, in der das Modell AUS DER ROLLE gefallen
|
||||
# ist (sich selbst als Claude bezeichnet, die ARIA-Persona als injiziert/erfunden
|
||||
# abtut, die Session als Fake bezeichnet). Solche Antworten duerfen NICHT in die
|
||||
# Conversation-History — ein einziger gespeicherter Break zieht bei schwachen
|
||||
# Folgeturns eine Kaskade nach sich (das Modell setzt seine eigene Ablehnung fort).
|
||||
#
|
||||
# BEWUSST nur STARKE, selbstreferenzielle Marker — nicht das blosse Wort
|
||||
# "Injection"/"injizier" (das nutzt ARIA in Security-/Pentest-Projekten voellig
|
||||
# legitim). Getroffen wird nur das Muster "ICH bin Claude / die Persona ist
|
||||
# erfunden / diese Session ist injiziert".
|
||||
_IDENTITY_BREAK = re.compile(
|
||||
r"ich\s+bin\s+(?:allerdings\s+|ja\s+|nach\s+wie\s+vor\s+|weiterhin\s+)*claude|"
|
||||
r"i'?m\s+(?:still\s+|actually\s+)?claude\s+code|i\s+am\s+claude\b|"
|
||||
r"erfundene[nr]?\s+(?:tool|persona|schemas)|fabricated\s+persona|"
|
||||
r"fabrizierte?\s+(?:persona|gespr|konversation)|fabricated\s+conversation|"
|
||||
r"fake[- ]persona|injizierte[rn]?\s+(?:system-?prompt|kontext|persona)|"
|
||||
r"injected\s+(?:system\s*prompt|persona|context)|"
|
||||
r"diese\s+session\s+enthält\s+(?:einen|eine)\b.{0,40}injizier|"
|
||||
r"this\s+session\s+(?:contains|has|keeps|repeatedly)\b.{0,40}(?:inject|fabricat|fake)|"
|
||||
r"nicht\s+real\s+in\s+dieser\s+(?:umgebung|session)|not\s+real\s+in\s+this",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
|
||||
|
||||
def looks_like_identity_break(text: str) -> bool:
|
||||
"""True, wenn eine ARIA-Antwort aus der Rolle gefallen ist. Fuer den
|
||||
Gift-Waechter im Agent (nicht persistieren + Retry)."""
|
||||
return bool(text and _IDENTITY_BREAK.search(text))
|
||||
|
||||
|
||||
def build_time_section() -> str:
|
||||
"""Aktueller Zeitstempel — damit ARIA Timer korrekt anlegen kann
|
||||
und Watcher-Conditions mit hour_of_day etc. einordenbar bleiben."""
|
||||
|
||||
+66
-24
@@ -30,7 +30,8 @@ CONFIG_PATH = os.environ.get("LOCAL_LLM_CONFIG", "/shared/config/local_llm.json"
|
||||
|
||||
ESCALATE_MARKER = "<<ESCALATE>>"
|
||||
|
||||
DEFAULT_CONFIG = {"enabled": False, "localOnly": False, "toolVariant": "slim"}
|
||||
DEFAULT_CONFIG = {"enabled": False, "localOnly": False,
|
||||
"toolVariant": "slim", "localLlmModel": "qwen3-8b"}
|
||||
|
||||
|
||||
def load_config() -> dict:
|
||||
@@ -42,6 +43,9 @@ def load_config() -> dict:
|
||||
"enabled": bool(data.get("enabled", False)),
|
||||
"localOnly": bool(data.get("localOnly", False)),
|
||||
"toolVariant": data.get("toolVariant", "slim") or "slim",
|
||||
# Welches lokale Modell llama-swap laden soll (B0.5). Muss zu einem
|
||||
# Key in xtts/llama-swap/config.yaml passen.
|
||||
"localLlmModel": (data.get("localLlmModel") or "qwen3-8b").strip(),
|
||||
}
|
||||
except (FileNotFoundError, json.JSONDecodeError):
|
||||
return dict(DEFAULT_CONFIG)
|
||||
@@ -57,15 +61,17 @@ def load_config() -> dict:
|
||||
# lokale Tier soll nur die klaren Plauder-Turns abgreifen; Fehlklassifikation
|
||||
# faengt zusaetzlich das <<ESCALATE>> im Modell selbst ab.
|
||||
|
||||
# Wortstämme, die auf Tool-/Aktionsbedarf hindeuten → nicht lokal (in B1a).
|
||||
# CLAUDE-ONLY-Themen → nicht lokal. Seit B1b hat das lokale Tier Werkzeuge
|
||||
# (web_search, memory_search, trigger_timer, Spotify), daher gehen Wetter, News,
|
||||
# Fakten, Timer, Musik, Gedaechtnis-Suche jetzt LOKAL. Nur was das lokale Tier
|
||||
# nicht kann bleibt hier: Bilder, Skills, Projekte, OAuth, Smart-Home (keine
|
||||
# Anbindung), Kalender/Mail (kein Tool).
|
||||
_TOOL_HINTS = re.compile(
|
||||
r"\b(wetter|temperatur|regne|timer|erinner|weck|licht|lampe|steckdose|"
|
||||
r"spotify|musik|lied|song|playlist|lauter|leiser|"
|
||||
r"bild|generier|male?\b|zeichne|foto|"
|
||||
r"merk dir|memory|gedächtnis|erinnere dich|"
|
||||
r"skill|trigger|projekt|oauth|spotify|kalender|termin|"
|
||||
r"such(e|st)?\b|google|internet|news|nachrichten|aktie|kurs|"
|
||||
r"maild?|email|nachricht schreiben|sende)\b",
|
||||
r"\b(bild|generier|male?\b|malen|zeichne|foto|"
|
||||
r"skill|projekt|oauth|"
|
||||
r"licht|lampe|steckdose|rollade|heizung|"
|
||||
r"kalender|termin|"
|
||||
r"maild?|e-?mail|nachricht schreiben)\b",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
|
||||
@@ -81,6 +87,17 @@ _HARD_HINTS = re.compile(
|
||||
_MAX_LEN_FOR_LOCAL = 220 # laengere Nachrichten = eher komplexe Aufgaben → Claude
|
||||
|
||||
|
||||
def _strip_leading_hint_blocks(text: str) -> str:
|
||||
"""Fuehrende `[ ... ]`-Hint-Bloecke (GPS, Barge-In von der Bridge) weg —
|
||||
sonst blaeht der Praefix die Laenge auf und verfaelscht die Heuristik."""
|
||||
s = (text or "").strip()
|
||||
prev = None
|
||||
while prev != s:
|
||||
prev = s
|
||||
s = re.sub(r"^\s*\[[^\]]*\]\s*", "", s)
|
||||
return s
|
||||
|
||||
|
||||
def should_try_local(user_message: str, cfg: dict) -> bool:
|
||||
"""True, wenn der Router diesen Turn (B1a, reden-only) lokal versuchen soll.
|
||||
localOnly überschreibt die Heuristik (dann IMMER lokal)."""
|
||||
@@ -88,7 +105,7 @@ def should_try_local(user_message: str, cfg: dict) -> bool:
|
||||
return False
|
||||
if cfg.get("localOnly"):
|
||||
return True
|
||||
msg = (user_message or "").strip()
|
||||
msg = _strip_leading_hint_blocks(user_message)
|
||||
if not msg or len(msg) > _MAX_LEN_FOR_LOCAL:
|
||||
return False
|
||||
if _TOOL_HINTS.search(msg):
|
||||
@@ -104,31 +121,56 @@ def should_try_local(user_message: str, cfg: dict) -> bool:
|
||||
# (WAS ARIA kann, ohne volle Schemas) + Escalation-Regel. KEINE Tool-Schemas,
|
||||
# kein volles Memory (B1a).
|
||||
|
||||
# Was das lokale Tier NICHT selbst kann → dafuer eskaliert es an Claude.
|
||||
_AWARENESS = (
|
||||
"ARIA kann ausserdem (ueber ihr grosses Backend, nicht in diesem schnellen "
|
||||
"Modus): Wetter/Zeit/Websuche, Timer & Erinnerungen, Musik (Spotify) & Licht "
|
||||
"steuern, Bilder generieren, ins Gedaechtnis schreiben/suchen, Skills bauen, "
|
||||
"Projekte & OAuth verwalten."
|
||||
"Nur das grosse Modell (nicht du im Schnell-Modus) kann: Bilder generieren, "
|
||||
"Skills bauen/aendern, Projekte & OAuth verwalten, ins Gedaechtnis SCHREIBEN, "
|
||||
"sowie tiefe/technische Analysen und langen Code."
|
||||
)
|
||||
|
||||
|
||||
def build_local_system_prompt(identity_anchor: str, pinned_persona: str = "") -> str:
|
||||
def build_local_system_prompt(identity_anchor: str, has_tools: bool = False,
|
||||
pinned_persona: str = "") -> str:
|
||||
"""Schlanker System-Prompt fuers lokale LLM. identity_anchor = derselbe
|
||||
Anker wie bei Claude (Rolle haelt); pinned_persona optional kurz."""
|
||||
Anker wie bei Claude (Rolle haelt)."""
|
||||
parts = [
|
||||
identity_anchor.strip(),
|
||||
"",
|
||||
"## SCHNELL-MODUS",
|
||||
"Du laeufst gerade als schnelles lokales Modell fuer einfache, kurze "
|
||||
"Alltags-Konversation. Antworte knapp, freundlich, auf Deutsch, als ARIA.",
|
||||
"Du laeufst gerade als schnelles lokales Modell fuer Alltags-Konversation "
|
||||
"und einfache Aufgaben. Antworte knapp, freundlich, auf Deutsch, als ARIA.",
|
||||
"WICHTIG zur Ausgabe: Antworte in ganz NORMALEM Text. Verwende KEINE "
|
||||
"`<voice>`-Tags, kein `[FILE:]`, kein `<tool_call>`, kein HTML/Markup — "
|
||||
"nur ein oder zwei natuerliche Saetze. (Der Text wird direkt angezeigt "
|
||||
"UND vorgelesen.)",
|
||||
]
|
||||
if has_tools:
|
||||
parts += [
|
||||
"",
|
||||
"## DEINE WERKZEUGE — nur nutzen wenn die Frage es WIRKLICH braucht",
|
||||
"- `web_search`: aktuelle Infos aus dem Netz — Wetter, News, Fakten, "
|
||||
"Preise, Oeffnungszeiten. Bei Wetter: nimm Stefans Ort aus dem "
|
||||
"GPS-Hinweis in der Nachricht.",
|
||||
"- `memory_search`: in ARIAs Gedaechtnis nachsehen (lesen).",
|
||||
"- `trigger_timer`: Timer/Erinnerung setzen ('in 10 Minuten…').",
|
||||
"- `run_spotify`: Musik steuern (naechster Titel, Pause, weiter).",
|
||||
"WICHTIG: Bei reinem Smalltalk ('wie gehts', Begruessung, Meinung) "
|
||||
"KEIN Werkzeug — einfach direkt antworten. Werkzeuge nur bei echtem "
|
||||
"Bedarf; erfinde keine.",
|
||||
]
|
||||
parts += [
|
||||
"",
|
||||
"## WAS DU HIER NICHT TUST",
|
||||
"## WAS DU HIER NICHT KANNST",
|
||||
_AWARENESS,
|
||||
"Fuer alles, was ein Werkzeug/eine Aktion braucht, tiefer/technischer wird, "
|
||||
"aktuelles Wissen oder Nachschlagen verlangt, oder wo du unsicher bist: "
|
||||
f"antworte AUSSCHLIESSLICH mit exakt `{ESCALATE_MARKER}` (nichts sonst). "
|
||||
"Dann uebernimmt automatisch das grosse Modell mit vollem Zugriff. "
|
||||
"Lieber einmal eskalieren als falsch raten.",
|
||||
"WICHTIG — du hast Stefans GEDAECHTNIS hier NICHT im Kopf: Bei Fragen zu "
|
||||
"seinem Leben, zu Personen/Namen, Beziehungen, seiner Vergangenheit, "
|
||||
"seinen Vorlieben/Sachen oder anderem gespeicherten Wissen antworte NICHT "
|
||||
"aus dem Nichts (und rate nicht, wer wer ist) — sondern eskaliere. Das "
|
||||
"grosse Modell kennt das Gedaechtnis und antwortet diskret.",
|
||||
"Dafuer — und bei tiefen/technischen Fragen, langem Code, oder wenn du "
|
||||
f"unsicher bist — antworte AUSSCHLIESSLICH mit exakt `{ESCALATE_MARKER}` "
|
||||
"(nichts sonst). Dann uebernimmt das grosse Modell mit vollem Zugriff. "
|
||||
"Lieber einmal eskalieren als falsch raten oder ein Werkzeug erfinden.",
|
||||
]
|
||||
if pinned_persona.strip():
|
||||
parts += ["", "## PERSONA", pinned_persona.strip()]
|
||||
|
||||
@@ -0,0 +1,25 @@
|
||||
# SearXNG-Config fuer ARIA (self-hosted Meta-Suche, Backend fuers web_search-Tool).
|
||||
# Erbt alle Default-Engines; wir ueberschreiben nur das Noetige:
|
||||
# - JSON-Format aktiviert (Default AUS) -> Brain kann /search?format=json rufen
|
||||
# - Rate-Limiter aus -> programmatischer Brain-Zugriff wird nicht geblockt
|
||||
# - eigener secret_key (interne Instanz auf aria-net, nicht oeffentlich exponiert)
|
||||
use_default_settings: true
|
||||
|
||||
server:
|
||||
# Interner Dienst auf aria-net, nicht oeffentlich. Trotzdem ein eigener Key.
|
||||
# Bei Bedarf aendern (beliebiger langer Zufallsstring).
|
||||
secret_key: "aria-searxng-6f2c9a1e8b7d4f30a5c1e2d9b8a7f6c3"
|
||||
limiter: false
|
||||
image_proxy: false
|
||||
|
||||
search:
|
||||
formats:
|
||||
- html
|
||||
- json
|
||||
# Deutsch bevorzugen (Brain kann per Query-Param ueberschreiben).
|
||||
default_lang: "de"
|
||||
|
||||
# Sanftere Timeouts, damit eine langsame Engine die Suche nicht ausbremst.
|
||||
outgoing:
|
||||
request_timeout: 5.0
|
||||
max_request_timeout: 10.0
|
||||
+119
-7
@@ -149,6 +149,54 @@ def _num_to_words_de(n: int) -> str:
|
||||
return str(n)
|
||||
|
||||
|
||||
# Vollstaendige Zehner fuer den generischen Konverter (inkl. 60-90).
|
||||
_TENS_FULL_DE = {20: "zwanzig", 30: "dreissig", 40: "vierzig", 50: "fuenfzig",
|
||||
60: "sechzig", 70: "siebzig", 80: "achtzig", 90: "neunzig"}
|
||||
|
||||
|
||||
def _below_100_de(n: int) -> str:
|
||||
if n in _NUM_WORDS_DE: # 0..20
|
||||
return _NUM_WORDS_DE[n]
|
||||
tens = (n // 10) * 10
|
||||
ones = n % 10
|
||||
if ones == 0:
|
||||
return _TENS_FULL_DE[tens]
|
||||
ones_word = "ein" if ones == 1 else _NUM_WORDS_DE[ones]
|
||||
return f"{ones_word}und{_TENS_FULL_DE[tens]}"
|
||||
|
||||
|
||||
def _below_1000_de(n: int) -> str:
|
||||
if n < 100:
|
||||
return _below_100_de(n)
|
||||
h, rest = divmod(n, 100)
|
||||
h_word = ("ein" if h == 1 else _NUM_WORDS_DE[h]) + "hundert"
|
||||
return h_word if rest == 0 else h_word + _below_100_de(rest)
|
||||
|
||||
|
||||
def _int_to_words_de(n: int) -> str:
|
||||
"""Ganzzahl 0..999999 als zusammenhaengendes deutsches Wort
|
||||
('dreiundzwanzig', 'einhundert', 'zweitausendsechsundzwanzig').
|
||||
Groesser (IDs/Codes) → als Ziffern lassen (der Aufrufer gated zusaetzlich)."""
|
||||
if n < 0:
|
||||
return "minus " + _int_to_words_de(-n)
|
||||
if n < 1000:
|
||||
return _below_1000_de(n)
|
||||
if n < 1_000_000:
|
||||
th, rest = divmod(n, 1000)
|
||||
th_word = ("ein" if th == 1 else _below_1000_de(th)) + "tausend"
|
||||
return th_word if rest == 0 else th_word + _below_1000_de(rest)
|
||||
return str(n)
|
||||
|
||||
|
||||
def _spell_standalone_int(m) -> str:
|
||||
"""Regex-Callback: freistehende Ganzzahl ausschreiben. Sehr lange
|
||||
Ziffernfolgen (IDs, Codes, Telefonnummern) bleiben Ziffern."""
|
||||
s = m.group(0)
|
||||
if len(s) > 6:
|
||||
return s
|
||||
return _int_to_words_de(int(s))
|
||||
|
||||
|
||||
def _time_range_to_words(m):
|
||||
"""'8:00-9:00 Uhr' → 'acht bis neun Uhr', '8-9 Uhr' → 'acht bis neun Uhr'."""
|
||||
h1 = int(m.group(1))
|
||||
@@ -168,7 +216,7 @@ def _decimal_to_words(m):
|
||||
"""'0.1' / '0,1' → 'null komma eins', '1,25' → 'eins komma zwei fuenf'."""
|
||||
int_part = int(m.group(1))
|
||||
dec_part = m.group(2)
|
||||
int_word = _num_to_words_de(int_part) if 0 <= int_part <= 59 else str(int_part)
|
||||
int_word = _int_to_words_de(int_part) if int_part <= 999999 else str(int_part)
|
||||
dec_words = " ".join(_num_to_words_de(int(d)) for d in dec_part)
|
||||
return f"{int_word} komma {dec_words}"
|
||||
|
||||
@@ -184,7 +232,7 @@ _UNIT_WORDS = [
|
||||
(r'\bkm\b', 'Kilometer'),
|
||||
(r'\bm/s\b', 'Meter pro Sekunde'),
|
||||
(r'\bkg\b', 'Kilogramm'),
|
||||
(r'\b°C\b', 'Grad Celsius'),
|
||||
(r'\b°C\b', ' Grad Celsius'),
|
||||
(r'°C', ' Grad Celsius'),
|
||||
(r'\bMbps\b', 'Megabit pro Sekunde'),
|
||||
(r'\bGbps\b', 'Gigabit pro Sekunde'),
|
||||
@@ -301,6 +349,15 @@ def clean_text_for_tts(text: str) -> str:
|
||||
# werden, koennen bei Bedarf explizit in _UNIT_WORDS uebersteuert werden.
|
||||
t = _re_tts.sub(r'\b([A-Z]{2,5})\b', lambda m: " ".join(m.group(1)), t)
|
||||
|
||||
# Generisches Ausschreiben ALLER verbleibenden freistehenden Ganzzahlen
|
||||
# ('23' → 'dreiundzwanzig', '100' → 'einhundert', '2026' → 'zwei-
|
||||
# tausendsechsundzwanzig'). Laeuft NACH Uhrzeiten/Dezimalzahlen/Einheiten,
|
||||
# die ihre Ziffern schon zu Woertern gemacht haben. Tag-unabhaengig — so
|
||||
# klingen auch Antworten des lokalen LLM (das keine <voice>-Tags nutzt)
|
||||
# sauber. Ziffernfolgen die an ., :, / kleben (IPs, Versionen, Uhrzeit-
|
||||
# Reste) ueberspringen wir, um sie nicht zu zerreissen.
|
||||
t = _re_tts.sub(r'(?<![\d.,:/])\d{1,6}(?![\d.,:/])', _spell_standalone_int, t)
|
||||
|
||||
# Anfuehrungszeichen
|
||||
t = _re_tts.sub(r'["""„`]', '', t)
|
||||
|
||||
@@ -1257,12 +1314,14 @@ class ARIABridge:
|
||||
# Voice-Tag-Noise als Kontext sieht).
|
||||
# File-Marker werden separat als file_from_aria-Events ausgeliefert.
|
||||
display_text = strip_voice_tag_for_display(text)
|
||||
_answered_by = (payload.get("answeredBy") or "") if isinstance(payload, dict) else ""
|
||||
assistant_backup_ts = self._append_chat_backup({
|
||||
"role": "assistant",
|
||||
"text": display_text,
|
||||
"files": [{"serverPath": f["serverPath"], "name": f["name"],
|
||||
"mimeType": f["mimeType"], "size": f["size"]} for f in aria_files],
|
||||
"project_id": turn_pid,
|
||||
"answeredBy": _answered_by,
|
||||
})
|
||||
|
||||
metadata = payload.get("metadata", {})
|
||||
@@ -1305,10 +1364,19 @@ class ARIABridge:
|
||||
# Projekt-Zuordnung — App + Diagnostic sortieren die Bubble in
|
||||
# den passenden Projekt-Block. Leer = Hauptchat.
|
||||
"projectId": (payload.get("projectId") or "") if isinstance(payload, dict) else "",
|
||||
# Quell-Backend (local/claude/fast-path) fuer den Diagnostic-Badge.
|
||||
"answeredBy": (payload.get("answeredBy") or "") if isinstance(payload, dict) else "",
|
||||
},
|
||||
"timestamp": int(asyncio.get_event_loop().time() * 1000),
|
||||
})
|
||||
|
||||
# System-Flag vom Brain: reine Steuerbefehle (Fast-Path) NICHT vorlesen.
|
||||
# Robust und unabhaengig von <voice>-Tags im Text (die ARIA beim
|
||||
# Skill-Rebuild verlieren kann) — die Quelle entscheidet, nicht der Inhalt.
|
||||
if isinstance(payload, dict) and payload.get("speak") is False:
|
||||
logger.info("[core] TTS uebersprungen (speak=False — Fast-Path-Steuerbefehl)")
|
||||
return
|
||||
|
||||
# TTS ueber XTTS (XTTS-Bridge auf Gaming-PC)
|
||||
if not (getattr(self, 'tts_enabled', True) and should_speak(self.current_mode, is_critical)):
|
||||
logger.info("[core] TTS unterdrueckt (Modus: %s)", self.current_mode.config.name)
|
||||
@@ -1633,6 +1701,12 @@ class ARIABridge:
|
||||
# gegeben damit der chat-Broadcast die Bubble dem richtigen Projekt-
|
||||
# Block in App + Diagnostic zuordnen kann.
|
||||
turn_project_id = (data.get("project_id") or "").strip()
|
||||
# Welcher Backend geantwortet hat (local/claude/fast-path) — fuer den
|
||||
# Quell-Badge in Diagnostic.
|
||||
answered_by = (data.get("answered_by") or "claude").strip()
|
||||
# Soll vorgelesen werden? Fast-Path (Steuerbefehl) = False. System-Flag
|
||||
# vom Brain — robust, unabhaengig von <voice>-Tags im Reply-Text.
|
||||
speak = data.get("speak", True)
|
||||
|
||||
# Side-Channel-Events VOR der Chat-Bubble broadcasten (z.B. skill_created)
|
||||
# damit sie in der UI vor der Reply auftauchen
|
||||
@@ -1699,7 +1773,9 @@ class ARIABridge:
|
||||
# passend behandelt wird (hier minimal, weil Brain noch keine
|
||||
# metadata mitschickt).
|
||||
try:
|
||||
await self._process_core_response(reply, {"projectId": turn_project_id})
|
||||
await self._process_core_response(reply, {"projectId": turn_project_id,
|
||||
"answeredBy": answered_by,
|
||||
"speak": speak})
|
||||
except Exception:
|
||||
logger.exception("[brain] _process_core_response Fehler")
|
||||
await self._emit_activity("idle", "", project_id=project_id)
|
||||
@@ -1710,6 +1786,14 @@ class ARIABridge:
|
||||
"timestamp": int(asyncio.get_event_loop().time() * 1000),
|
||||
})
|
||||
|
||||
# Das 'thinking' oben ging mit der REQUEST-projectId raus, das 'idle' in
|
||||
# _process_core_response aber mit der TURN-projectId (Brain kann den Turn
|
||||
# umgeroutet haben, z.B. Voice-Sticky). Weichen sie ab, bekaeme der
|
||||
# Request-Kontext nie sein idle → "ARIA denkt" bliebe dort haengen.
|
||||
# Darum hier zusaetzlich fuer die Request-projectId ein idle.
|
||||
if (project_id or "") != (turn_project_id or ""):
|
||||
await self._emit_activity("idle", "", project_id=project_id)
|
||||
|
||||
if data.get("distilling"):
|
||||
logger.info("[brain] Destillat laeuft im Hintergrund")
|
||||
|
||||
@@ -2922,6 +3006,24 @@ class ARIABridge:
|
||||
if self._is_duplicate_client_msg(client_msg_id):
|
||||
return
|
||||
|
||||
# Text-Fenster-Dedup: im Auto cancelt der App-No-Speech-Watchdog den
|
||||
# Stream vorzeitig und startet passives Lauschen, waehrend die Bridge
|
||||
# dieselbe Aeusserung parallel noch finalisiert → derselbe Befehl
|
||||
# kaeme zweimal (aus ZWEI verschiedenen audioRequestIds, darum greift
|
||||
# die ID-Idempotenz oben nicht). Identischen Endpoint-Text innerhalb
|
||||
# 5s verwerfen. Genuines "nächstes, nächstes" liegt praktisch immer
|
||||
# weiter auseinander (man hoert den Wechsel erst).
|
||||
_norm_txt = " ".join(text.lower().split())
|
||||
_now_t = asyncio.get_event_loop().time()
|
||||
if (_norm_txt
|
||||
and _norm_txt == getattr(self, "_last_endpoint_norm", "")
|
||||
and (_now_t - getattr(self, "_last_endpoint_at", 0.0)) < 5.0):
|
||||
logger.info("[rvs] stt_endpoint Dupe verworfen (gleicher Text <5s): %r",
|
||||
text[:60])
|
||||
return
|
||||
self._last_endpoint_norm = _norm_txt
|
||||
self._last_endpoint_at = _now_t
|
||||
|
||||
# App-Focus aus stt_stream_start-Registry auflösen (falls die
|
||||
# App-Version die projectId noch nicht mitschickt: leer = Hauptchat).
|
||||
stream_req_id = payload.get("requestId", "") or ""
|
||||
@@ -3367,9 +3469,11 @@ class ARIABridge:
|
||||
_LLM_TIMEOUT_S = 30.0
|
||||
|
||||
async def _local_llm(self, messages: list, max_tokens: int = 512,
|
||||
temperature: float = 0.7, stop=None) -> dict:
|
||||
temperature: float = 0.7, stop=None, tools=None,
|
||||
model=None) -> dict:
|
||||
"""Schickt einen llm_request an den llm-adapter (Gamebox), wartet auf
|
||||
llm_response. Rueckgabe: {ok, content, model, elapsedMs} oder {ok:False, error}."""
|
||||
llm_response. tools (B1b) werden durchgereicht; tool_calls kommen zurueck.
|
||||
Rueckgabe: {ok, content, tool_calls, model, elapsedMs} oder {ok:False, error}."""
|
||||
if self.ws_rvs is None:
|
||||
return {"ok": False, "error": "RVS-Verbindung nicht aktiv"}
|
||||
if not isinstance(messages, list) or not messages:
|
||||
@@ -3388,8 +3492,12 @@ class ARIABridge:
|
||||
}
|
||||
if stop:
|
||||
req_payload["stop"] = stop
|
||||
logger.info("[rvs] llm_request → llm-adapter (id=%s, msgs=%d, max_tokens=%d)",
|
||||
request_id[:8], len(messages), max_tokens)
|
||||
if tools:
|
||||
req_payload["tools"] = tools
|
||||
if model:
|
||||
req_payload["model"] = model
|
||||
logger.info("[rvs] llm_request → llm-adapter (id=%s, msgs=%d, max_tokens=%d, tools=%d, model=%s)",
|
||||
request_id[:8], len(messages), max_tokens, len(tools) if tools else 0, model or "-")
|
||||
ok = await self._send_to_rvs({
|
||||
"type": "llm_request",
|
||||
"payload": req_payload,
|
||||
@@ -3407,6 +3515,7 @@ class ARIABridge:
|
||||
return {
|
||||
"ok": True,
|
||||
"content": result.get("content", ""),
|
||||
"tool_calls": result.get("tool_calls"),
|
||||
"model": result.get("model"),
|
||||
"elapsedMs": result.get("elapsedMs"),
|
||||
}
|
||||
@@ -3885,9 +3994,12 @@ class ARIABridge:
|
||||
temperature = float(data.get("temperature"))
|
||||
except (TypeError, ValueError):
|
||||
temperature = 0.7
|
||||
_tools = data.get("tools") if isinstance(data.get("tools"), list) else None
|
||||
_model = data.get("model") if isinstance(data.get("model"), str) else None
|
||||
result = await self._local_llm(
|
||||
messages=messages, max_tokens=max_tokens,
|
||||
temperature=temperature, stop=data.get("stop"),
|
||||
tools=_tools, model=_model,
|
||||
)
|
||||
status = 200 if result.get("ok") else 502
|
||||
await _send_response(writer, status, result)
|
||||
|
||||
+266
-32
@@ -912,30 +912,66 @@
|
||||
|
||||
<div class="settings-section">
|
||||
<h2>Lokales LLM (schnelle Antworten) <button class="info-btn" onclick="showInfo('local-llm')" title="Wie funktioniert das?">ℹ</button></h2>
|
||||
<div class="card" style="max-width:500px;">
|
||||
<div style="font-size:11px;color:#8888AA;margin-bottom:10px;line-height:1.5;">
|
||||
Ein schnelles lokales Modell (Qwen3 auf der Gamebox) beantwortet einfache
|
||||
Plauder-Turns in unter 1 Sekunde; alles Schwere/Werkzeug-artige geht
|
||||
weiter an Claude. Schreibt <code>/shared/config/local_llm.json</code> —
|
||||
der Router im Brain liest sie live (kein Neustart noetig).
|
||||
<div class="card" style="max-width:540px;">
|
||||
<div style="font-size:11px;color:#8888AA;margin-bottom:12px;line-height:1.55;">
|
||||
Ein schnelles lokales Modell (<strong>Qwen3</strong> auf der Gamebox) beantwortet
|
||||
<strong>leichte Fragen in unter 1 Sekunde</strong>. Was schwerer wird, gibt es
|
||||
automatisch an <strong>Claude</strong> ab. Aenderungen greifen sofort
|
||||
(kein Neustart) — geschrieben nach <code>/shared/config/local_llm.json</code>.
|
||||
</div>
|
||||
<label style="display:flex;align-items:center;gap:8px;margin-bottom:8px;font-size:12px;color:#E0E0F0;cursor:pointer;">
|
||||
<input type="checkbox" id="local-llm-enabled" onchange="saveLocalLlmConfig()" style="margin:0;">
|
||||
<span>Lokales LLM nutzen <span style="color:#8888AA;">(aus = alles ueber Claude)</span></span>
|
||||
|
||||
<!-- Master -->
|
||||
<label style="display:flex;align-items:flex-start;gap:8px;margin-bottom:4px;font-size:13px;color:#E0E0F0;cursor:pointer;">
|
||||
<input type="checkbox" id="local-llm-enabled" onchange="saveLocalLlmConfig()" style="margin:2px 0 0 0;">
|
||||
<span><strong>Lokales LLM einschalten</strong></span>
|
||||
</label>
|
||||
<label style="display:flex;align-items:center;gap:8px;margin-bottom:10px;font-size:12px;color:#E0E0F0;cursor:pointer;">
|
||||
<input type="checkbox" id="local-llm-onlylocal" onchange="saveLocalLlmConfig()" style="margin:0;">
|
||||
<span>Nur lokales LLM — kein Claude-Fallback <span style="color:#8888AA;">(zum Testen)</span></span>
|
||||
<div style="font-size:10px;color:#8888AA;margin:0 0 12px 24px;line-height:1.5;">
|
||||
<strong style="color:#4ADE80;">AN:</strong> leichte Fragen → lokal (schnell), schwere → Claude.
|
||||
<strong style="color:#8888AA;">AUS:</strong> alles laeuft ueber Claude (wie bisher). — <em>Das ist der Normal-Betrieb.</em>
|
||||
</div>
|
||||
|
||||
<!-- Nur lokal -->
|
||||
<label style="display:flex;align-items:flex-start;gap:8px;margin-bottom:4px;font-size:13px;color:#E0E0F0;cursor:pointer;">
|
||||
<input type="checkbox" id="local-llm-onlylocal" onchange="saveLocalLlmConfig()" style="margin:2px 0 0 0;">
|
||||
<span><strong>Nur lokales LLM</strong> — Claude komplett aussperren</span>
|
||||
</label>
|
||||
<div style="display:flex;align-items:center;gap:8px;margin-bottom:6px;">
|
||||
<span style="font-size:12px;color:#8888AA;">Tool-Umfang:</span>
|
||||
<div style="font-size:10px;color:#8888AA;margin:0 0 12px 24px;line-height:1.5;">
|
||||
<strong style="color:#FFD60A;">Nur zum Testen.</strong> Erzwingt IMMER das lokale Modell — auch bei
|
||||
schweren Fragen, ohne Claude-Rettung. So siehst du, was Qwen allein kann. Werkzeug-Fragen
|
||||
(Wetter/Timer/…) funktionieren dann nicht. <strong>Fuer den Alltag: AUS lassen.</strong>
|
||||
</div>
|
||||
|
||||
<!-- Tool-Umfang -->
|
||||
<div style="display:flex;align-items:center;gap:8px;margin-bottom:4px;">
|
||||
<span style="font-size:13px;color:#E0E0F0;"><strong>Werkzeuge lokal:</strong></span>
|
||||
<select id="local-llm-toolvariant" onchange="saveLocalLlmConfig()" style="background:#1E1E2E;border:1px solid #333;border-radius:4px;padding:6px 8px;color:#E0E0F0;font-family:inherit;font-size:12px;">
|
||||
<option value="slim">Abgespeckt (kuratierte Tools)</option>
|
||||
<option value="full" disabled>Voll (braucht mehr VRAM)</option>
|
||||
<option value="slim">Abgespeckt — kuratierte Tools</option>
|
||||
<option value="full" disabled>Voll — ganzes Arsenal (braucht mehr VRAM)</option>
|
||||
</select>
|
||||
<button class="info-btn" onclick="showInfo('local-llm-tools')" title="Voll: wie viel VRAM?">ℹ</button>
|
||||
</div>
|
||||
<div id="local-llm-status" style="font-size:10px;color:#6a6a88;margin-top:6px;min-height:12px;"></div>
|
||||
<div style="font-size:10px;color:#8888AA;margin:0 0 6px 0;line-height:1.5;">
|
||||
Welche Werkzeuge das lokale Modell <em>selbst</em> ausfuehren darf. „Voll" ist gesperrt,
|
||||
bis eine 2. Grafikkarte da ist (siehe ⓘ).
|
||||
<br><span style="color:#FFD60A;">Aktueller Stand (B1a): das lokale Modell <strong>plaudert nur</strong> — Werkzeuge macht noch Claude. Lokale Tools kommen mit B1b.</span>
|
||||
</div>
|
||||
|
||||
<!-- Lokales Modell (llama-swap, B0.5) -->
|
||||
<div style="display:flex;align-items:center;gap:8px;margin:12px 0 4px 0;padding-top:10px;border-top:1px solid #2a2a3a;">
|
||||
<span style="font-size:13px;color:#E0E0F0;"><strong>Lokales Modell:</strong></span>
|
||||
<select id="local-llm-model" onchange="saveLocalLlmConfig()" style="flex:1;background:#1E1E2E;border:1px solid #333;border-radius:4px;padding:6px 8px;color:#E0E0F0;font-family:inherit;font-size:12px;">
|
||||
<option value="">(lade Liste…)</option>
|
||||
</select>
|
||||
<button class="btn secondary" onclick="loadLocalModelList()" title="Liste neu laden" style="padding:4px 8px;font-size:10px;">↻</button>
|
||||
</div>
|
||||
<div id="local-llm-model-desc" style="font-size:10px;color:#8888AA;margin:0 0 2px 0;line-height:1.5;"></div>
|
||||
<div style="font-size:10px;color:#FFD60A;margin:0 0 4px 0;line-height:1.5;">
|
||||
Beim ersten Wechsel zu einem Modell lädt die Gamebox das GGUF (mehrere GB) —
|
||||
die <strong>erste Antwort dauert dann länger</strong>, danach ist es gecacht.
|
||||
Liste kommt aus <code>/shared/config/local_models.json</code> (Keys = xtts/llama-swap/config.yaml).
|
||||
</div>
|
||||
|
||||
<div id="local-llm-status" style="font-size:11px;color:#6a6a88;margin-top:8px;padding-top:8px;border-top:1px solid #2a2a3a;min-height:14px;"></div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
@@ -992,6 +1028,18 @@
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="settings-section">
|
||||
<h2>Lokales LLM & Claude-Ersparnis <button class="info-btn" onclick="showInfo('local-savings')" title="Wie wird die Ersparnis gerechnet?">ℹ</button></h2>
|
||||
<div class="card">
|
||||
<div style="font-size:11px;color:#8888AA;margin-bottom:10px;">
|
||||
Turns, die das <strong style="color:#B392F0;">lokale LLM</strong> (Qwen) oder ein reiner <strong style="color:#F0B85E;">Skill-Fast-Path</strong> uebernommen hat — jeder davon ist ein Claude-Call, der NICHT passiert ist. Die lokalen Tokens laufen auf deiner eigenen Hardware (kein Subscription-Quota).
|
||||
</div>
|
||||
<div id="savings-grid" style="display:grid;grid-template-columns:repeat(auto-fit,minmax(150px,1fr));gap:8px;font-size:12px;">
|
||||
<div class="metric-cell"><div class="metric-label">–</div><div class="metric-value">–</div></div>
|
||||
</div>
|
||||
</div>
|
||||
</div>
|
||||
|
||||
<div class="settings-section">
|
||||
<h2>Bootstrap & Migration <button class="info-btn" onclick="showInfo('bootstrap')" title="Was sind die drei Wege?">ℹ</button></h2>
|
||||
<div class="card" style="line-height:1.6;">
|
||||
@@ -1047,6 +1095,7 @@
|
||||
<div style="display:flex;justify-content:space-between;align-items:center;margin-bottom:8px;">
|
||||
<h2 style="margin:0;">📁 Projekte</h2>
|
||||
<div>
|
||||
<button class="btn secondary" id="toggle-hidden-btn" onclick="toggleShowHidden()" style="padding:4px 10px;font-size:11px;" title="Versteckte Projekte ein-/ausblenden">👁 Versteckte anzeigen</button>
|
||||
<button class="btn secondary" onclick="loadProjects()" style="padding:4px 10px;font-size:11px;">🔄 Aktualisieren</button>
|
||||
<button class="btn" onclick="openCreateProjectModal()" style="padding:4px 10px;font-size:11px;">+ Neues Projekt</button>
|
||||
</div>
|
||||
@@ -1375,6 +1424,7 @@
|
||||
let focusedContextId = localStorage.getItem('diag_focused_context_id') || '';
|
||||
let diagQueueStatus = {};
|
||||
let diagProjectsCache = [];
|
||||
let diagShowHiddenStrip = false; // versteckte Projekte im Streifen zeigen?
|
||||
|
||||
function updateChatVisibilityByFocus() {
|
||||
for (const box of [chatBox, document.getElementById('chat-box-fs')]) {
|
||||
@@ -1415,20 +1465,79 @@
|
||||
if (s.queue_size > 0) return { color: '#FFD60A', label: `Queue: ${s.queue_size}` };
|
||||
return { color: '#34C759', label: 'idle' };
|
||||
};
|
||||
// Projekt-Chip MIT Auge (verstecken/sichtbar). Auge hat eigenes onclick
|
||||
// + stopPropagation, damit der Klick nicht den Kontext wechselt.
|
||||
const projChip = (p, isFocus, dotColor, subline) => {
|
||||
const hidden = !!p.hidden;
|
||||
const bg = isFocus ? 'rgba(52,199,89,0.15)' : '#1E1E2E';
|
||||
const border = isFocus ? '#34C759' : '#2A2A3E';
|
||||
const eye = hidden ? '👁' : '🙈';
|
||||
const eyeTitle = hidden ? 'Wieder dauerhaft sichtbar machen' : 'Verstecken (aus dem Streifen ausblenden)';
|
||||
return `<div style="flex:0 0 auto;padding:6px 10px;background:${bg};border:1px solid ${border};border-radius:6px;display:flex;align-items:center;gap:6px;min-width:120px;${hidden ? 'opacity:0.5;' : ''}">
|
||||
<div onclick="switchDiagFocus('${p.id}')" style="cursor:pointer;display:flex;align-items:center;gap:6px;min-width:0;">
|
||||
<div style="width:8px;height:8px;border-radius:4px;background:${dotColor};"></div>
|
||||
<div style="display:flex;flex-direction:column;min-width:0;">
|
||||
<div style="color:${isFocus?'#34C759':'#E0E0F0'};font-size:12px;font-weight:600;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;max-width:200px;">📁 ${escapeHtml(p.name)}${hidden ? ' <span style="color:#B392F0;font-size:9px;font-weight:700;">versteckt</span>' : ''}</div>
|
||||
<div style="color:#8888AA;font-size:10px;">${subline}</div>
|
||||
</div>
|
||||
</div>
|
||||
<span onclick="event.stopPropagation();setStripProjectHidden('${p.id}',${!hidden})" title="${eyeTitle}" style="cursor:pointer;font-size:14px;padding:2px 4px;user-select:none;">${eye}</span>
|
||||
</div>`;
|
||||
};
|
||||
const cards = [];
|
||||
// Hauptchat
|
||||
const mainDot = dotFor('__main__');
|
||||
cards.push(chip('', '💬 Hauptchat', focusedContextId === '', mainDot.color, mainDot.label || 'idle'));
|
||||
// Projekte — nur active/ended, sortiert nach letzter Aktivitaet
|
||||
// Projekte — nur active/ended, sortiert nach letzter Aktivitaet.
|
||||
// Versteckte standardmaessig raus; per Toggle-Chip einblendbar.
|
||||
let hiddenCount = 0;
|
||||
for (const p of diagProjectsCache) {
|
||||
if (p.status === 'archived') continue;
|
||||
if (p.hidden) {
|
||||
hiddenCount++;
|
||||
if (!diagShowHiddenStrip) continue;
|
||||
}
|
||||
const d = dotFor(p.id);
|
||||
const sub = d.label || `${p.turn_count} Turns`;
|
||||
cards.push(chip(p.id, `📁 ${p.name}`, focusedContextId === p.id, d.color, sub));
|
||||
cards.push(projChip(p, focusedContextId === p.id, d.color, sub));
|
||||
}
|
||||
// Toggle-Chip am Ende (nur wenn es versteckte gibt oder gerade gezeigt werden)
|
||||
if (hiddenCount > 0 || diagShowHiddenStrip) {
|
||||
const tLabel = diagShowHiddenStrip
|
||||
? `🙈 versteckte ausblenden (${hiddenCount})`
|
||||
: `👁 versteckte anzeigen (${hiddenCount})`;
|
||||
cards.push(`<div onclick="toggleDiagHiddenStrip()" title="Versteckte Projekte ein-/ausblenden" style="cursor:pointer;flex:0 0 auto;padding:6px 10px;background:#0D0D18;border:1px dashed #3A3A50;border-radius:6px;display:flex;align-items:center;color:#B392F0;font-size:11px;font-weight:600;white-space:nowrap;">${tLabel}</div>`);
|
||||
}
|
||||
strip.innerHTML = cards.join('');
|
||||
}
|
||||
|
||||
function toggleDiagHiddenStrip() {
|
||||
diagShowHiddenStrip = !diagShowHiddenStrip;
|
||||
renderContextStrip();
|
||||
}
|
||||
|
||||
async function setStripProjectHidden(id, hidden) {
|
||||
try {
|
||||
const r = await fetch(`/api/brain/projects/${encodeURIComponent(id)}`, {
|
||||
method: 'PATCH',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({ hidden: !!hidden }),
|
||||
});
|
||||
if (!r.ok) throw new Error('HTTP ' + r.status);
|
||||
// Wenn wir das gerade fokussierte Projekt verstecken (und Versteckte nicht
|
||||
// eingeblendet sind), zurueck auf Hauptchat — sonst haengt der Focus an
|
||||
// einem unsichtbaren Chip.
|
||||
if (hidden && id === focusedContextId && !diagShowHiddenStrip) {
|
||||
switchDiagFocus('');
|
||||
}
|
||||
// Lokal sofort aktualisieren (der /api/brain-Proxy broadcastet zusaetzlich
|
||||
// project_changed an App + andere Tabs).
|
||||
await refreshDiagProjectsCache();
|
||||
} catch (e) {
|
||||
alert('Verstecken/Anzeigen fehlgeschlagen: ' + e.message);
|
||||
}
|
||||
}
|
||||
|
||||
async function refreshDiagQueueStatus() {
|
||||
try {
|
||||
const r = await fetch('/api/brain/projects/queue-status');
|
||||
@@ -1550,8 +1659,8 @@
|
||||
try { loadBrainStatus(); } catch {}
|
||||
// Sprachmodell-Dropdown befuellen (kuratierte Tier-Liste vom Proxy)
|
||||
try { loadModelList(); } catch {}
|
||||
// Lokales-LLM-Schalter aus /shared/config/local_llm.json laden
|
||||
try { loadLocalLlmConfig(); } catch {}
|
||||
// Lokales-LLM: erst Modell-Liste (Dropdown), dann Config (Auswahl setzen)
|
||||
try { loadLocalModelList().then(() => loadLocalLlmConfig()); } catch {}
|
||||
};
|
||||
|
||||
// Brain-Status periodisch refreshen damit die Card live bleibt
|
||||
@@ -1726,8 +1835,10 @@
|
||||
}
|
||||
|
||||
if (msg.type === 'project_changed') {
|
||||
// ARIA hat in einem Tool-Call ein Projekt erstellt/betreten/verlassen/beendet.
|
||||
// Liste neu laden falls sichtbar.
|
||||
// Projekt geaendert (ARIA-Tool, App/Diagnostic-Verstecken, …) →
|
||||
// BEIDE Projekt-UIs live aktualisieren: den Kontext-Streifen (Main)
|
||||
// und die vertikale Liste (Einstellungen).
|
||||
refreshDiagProjectsCache();
|
||||
loadProjects();
|
||||
return;
|
||||
}
|
||||
@@ -1870,6 +1981,7 @@
|
||||
ttsText: p.ttsText,
|
||||
backupTs: p.backupTs,
|
||||
projectId: p.projectId || '',
|
||||
answeredBy: p.answeredBy || '',
|
||||
});
|
||||
return;
|
||||
}
|
||||
@@ -1971,7 +2083,8 @@
|
||||
const trashBtn = m.ts
|
||||
? `<button class="bubble-trash" title="Diese Bubble loeschen" onclick="deleteDiagBubble(${m.ts})">🗑</button>`
|
||||
: '';
|
||||
const innerHtml = `${trashBtn}${linked}<div class="meta">${escapeHtml(m.meta)} — ${time}</div>`;
|
||||
const histBadge = srcBadgeHtml(m.type, m.answeredBy || '');
|
||||
const innerHtml = `${trashBtn}${linked}<div class="meta">${escapeHtml(m.meta)}${histBadge} — ${time}</div>`;
|
||||
for (const b of boxes) {
|
||||
const el = document.createElement('div');
|
||||
el.className = `chat-msg ${m.type}`;
|
||||
@@ -2302,6 +2415,18 @@
|
||||
return t.trim();
|
||||
}
|
||||
|
||||
// Quell-Badge (local/claude/fast-path) fuer ARIA-Bubbles — in Live + History genutzt.
|
||||
function srcBadgeHtml(type, answeredBy) {
|
||||
if (type !== 'received' || !answeredBy) return '';
|
||||
const M = {
|
||||
'local': { t: '⚡ lokal', c: '#34C759' },
|
||||
'claude': { t: 'Claude', c: '#0096FF' },
|
||||
'fast-path': { t: '⚡ Fast-Path', c: '#AF7BFF' },
|
||||
};
|
||||
const b = M[answeredBy] || { t: answeredBy, c: '#8888AA' };
|
||||
return `<span title="Antwort erzeugt von: ${escapeHtml(answeredBy)}" style="display:inline-block;margin-left:6px;padding:1px 6px;border-radius:8px;font-size:9px;font-weight:bold;background:${b.c}22;color:${b.c};border:1px solid ${b.c}55;">${b.t}</span>`;
|
||||
}
|
||||
|
||||
function addChat(type, text, meta, options) {
|
||||
// [FILE: /shared/uploads/aria_xxx.ext]-Marker aus dem Antworttext entfernen —
|
||||
// die Datei kommt separat via file_from_aria-Event als eigene Bubble.
|
||||
@@ -2336,7 +2461,9 @@
|
||||
const trashBtn = backupTs
|
||||
? `<button class="bubble-trash" title="Diese Bubble loeschen" onclick="deleteDiagBubble(${backupTs})">🗑</button>`
|
||||
: '';
|
||||
const html = `${trashBtn}${linked}${ttsBlock}${gpsBlock}<div class="meta">${escapeHtml(meta)} — ${new Date().toLocaleTimeString('de-DE')}</div>`;
|
||||
// Quell-Badge: welcher Backend die Antwort erzeugt hat (nur ARIA-Bubbles)
|
||||
const srcBadge = srcBadgeHtml(type, (options && options.answeredBy) || '');
|
||||
const html = `${trashBtn}${linked}${ttsBlock}${gpsBlock}<div class="meta">${escapeHtml(meta)}${srcBadge} — ${new Date().toLocaleTimeString('de-DE')}</div>`;
|
||||
|
||||
// Thinking-Indikator ausblenden bei neuer Nachricht
|
||||
updateThinkingIndicator({ activity: 'idle' });
|
||||
@@ -2589,7 +2716,7 @@
|
||||
// Liste neu aufbauen
|
||||
list.innerHTML = '';
|
||||
let anyLoading = false, anyError = false;
|
||||
const labels = { f5tts: 'F5-TTS', whisper: 'Whisper STT', flux: 'FLUX Image-Gen' };
|
||||
const labels = { f5tts: 'F5-TTS', whisper: 'Whisper STT', flux: 'FLUX Image-Gen', llm: 'Lokales LLM' };
|
||||
for (const [s, info] of Object.entries(_serviceState)) {
|
||||
const row = document.createElement('div');
|
||||
row.style.cssText = 'display:flex;align-items:center;gap:6px;';
|
||||
@@ -2938,16 +3065,28 @@
|
||||
<div style="color:${!activeId ? '#34C759' : '#E0E0F0'};font-weight:600;">💬 Hauptchat ${!activeId ? '<span style="font-size:10px;font-weight:800;">✓ AKTIV</span>' : ''}</div>
|
||||
<div style="color:#555570;font-size:11px;margin-top:2px;">Standard-Verlauf, keine Projekt-Zuordnung</div>
|
||||
</div>`);
|
||||
const showHidden = !!window.__showHidden;
|
||||
let hiddenCount = 0, shownCount = 0;
|
||||
for (const p of projects) {
|
||||
const hidden = !!p.hidden;
|
||||
if (hidden) hiddenCount++;
|
||||
// Versteckte nur zeigen wenn der Toggle an ist.
|
||||
if (hidden && !showHidden) continue;
|
||||
shownCount++;
|
||||
const isActive = p.id === activeId;
|
||||
const since = p.last_activity_at ? new Date(p.last_activity_at * 1000).toLocaleString('de-DE') : '?';
|
||||
const ended = p.status === 'ended';
|
||||
// Auge: versteckt → 👁 (wieder sichtbar machen), sichtbar → 🙈 (verstecken).
|
||||
const eyeIcon = hidden ? '👁' : '🙈';
|
||||
const eyeTitle = hidden ? 'Wieder dauerhaft sichtbar machen' : 'Verstecken (aus Listen ausblenden)';
|
||||
const eyeColor = hidden ? '#B392F0' : '#8888AA';
|
||||
rows.push(`
|
||||
<div style="padding:12px 14px;border-bottom:1px solid #1E1E2E;${isActive ? 'background:rgba(52,199,89,0.08);border-left:3px solid #34C759;' : ''}">
|
||||
<div style="padding:12px 14px;border-bottom:1px solid #1E1E2E;${isActive ? 'background:rgba(52,199,89,0.08);border-left:3px solid #34C759;' : ''}${hidden ? 'opacity:0.55;' : ''}">
|
||||
<div style="display:flex;justify-content:space-between;align-items:flex-start;gap:8px;">
|
||||
<div onclick="switchProject('${p.id}')" style="cursor:pointer;flex:1;">
|
||||
<div style="color:${isActive ? '#34C759' : '#E0E0F0'};font-weight:600;">
|
||||
📁 ${escapeHtml(p.name)}
|
||||
${hidden ? '🙈' : '📁'} ${escapeHtml(p.name)}
|
||||
${hidden ? '<span style="color:#B392F0;font-size:10px;font-weight:700;margin-left:6px;background:rgba(179,146,240,0.15);padding:2px 6px;border-radius:3px;">versteckt</span>' : ''}
|
||||
${ended ? '<span style="color:#FFD60A;font-size:10px;font-weight:700;margin-left:6px;background:rgba(255,214,10,0.15);padding:2px 6px;border-radius:3px;">beendet</span>' : ''}
|
||||
${isActive ? '<span style="color:#34C759;font-size:10px;font-weight:800;margin-left:6px;">✓ AKTIV</span>' : ''}
|
||||
</div>
|
||||
@@ -2955,6 +3094,7 @@
|
||||
<div style="color:#555570;font-size:11px;margin-top:4px;">${p.turn_count} Turns · zuletzt ${since}</div>
|
||||
</div>
|
||||
<div style="display:flex;gap:4px;">
|
||||
<button class="btn secondary" onclick="setProjectHidden('${p.id}', ${!hidden})" style="padding:3px 8px;font-size:10px;color:${eyeColor};" title="${eyeTitle}">${eyeIcon}</button>
|
||||
${!ended ? `<button class="btn secondary" onclick="endProject('${p.id}', '${escapeHtmlAttr(p.name)}')" style="padding:3px 8px;font-size:10px;" title="Projekt beenden">⏹</button>` : ''}
|
||||
<button class="btn secondary" onclick="archiveProject('${p.id}', '${escapeHtmlAttr(p.name)}')" style="padding:3px 8px;font-size:10px;color:#E55C5C;" title="Archivieren">🗑</button>
|
||||
</div>
|
||||
@@ -2963,13 +3103,43 @@
|
||||
}
|
||||
if (projects.length === 0) {
|
||||
rows.push('<div style="padding:18px;color:#555570;font-size:12px;text-align:center;">Noch keine Projekte. „+ Neues Projekt" oder sag ARIA „lass uns ein Projekt anlegen".</div>');
|
||||
} else if (shownCount === 0) {
|
||||
rows.push(`<div style="padding:18px;color:#555570;font-size:12px;text-align:center;">Alle ${hiddenCount} Projekte sind versteckt. Klick „👁 Versteckte anzeigen".</div>`);
|
||||
}
|
||||
listEl.innerHTML = rows.join('');
|
||||
|
||||
// Toggle-Button beschriften (mit Anzahl) + Zustand spiegeln.
|
||||
const thBtn = document.getElementById('toggle-hidden-btn');
|
||||
if (thBtn) {
|
||||
thBtn.textContent = showHidden
|
||||
? `🙈 Versteckte ausblenden${hiddenCount ? ' ('+hiddenCount+')' : ''}`
|
||||
: `👁 Versteckte anzeigen${hiddenCount ? ' ('+hiddenCount+')' : ''}`;
|
||||
thBtn.style.opacity = hiddenCount ? '1' : '0.5';
|
||||
}
|
||||
} catch (e) {
|
||||
listEl.innerHTML = `<div style="padding:14px;color:#FF6E6E;font-size:12px;">Fehler: ${e.message}</div>`;
|
||||
}
|
||||
}
|
||||
|
||||
function toggleShowHidden() {
|
||||
window.__showHidden = !window.__showHidden;
|
||||
loadProjects();
|
||||
}
|
||||
|
||||
async function setProjectHidden(id, hidden) {
|
||||
try {
|
||||
const r = await fetch(`/api/brain/projects/${encodeURIComponent(id)}`, {
|
||||
method: 'PATCH',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({ hidden: !!hidden }),
|
||||
});
|
||||
if (!r.ok) throw new Error('HTTP ' + r.status);
|
||||
// Beim Verstecken bleibt der „anzeigen"-Modus wie er ist; beim
|
||||
// Wieder-Sichtbarmachen sieht man es sofort ohne Umschalten.
|
||||
loadProjects();
|
||||
} catch (e) { alert('Verstecken/Anzeigen fehlgeschlagen: ' + e.message); }
|
||||
}
|
||||
|
||||
async function switchProject(projectId) {
|
||||
try {
|
||||
await fetch('/api/brain/projects/switch', {
|
||||
@@ -5845,6 +6015,29 @@
|
||||
setCell('metrics-h24', d.h24);
|
||||
setCell('metrics-d30', d.d30);
|
||||
|
||||
// Lokales LLM & Claude-Ersparnis — pro Fenster: wie viele Claude-Calls
|
||||
// durch local/fast-path vermieden wurden + lokale Token-Last.
|
||||
const savingsGrid = document.getElementById('savings-grid');
|
||||
if (savingsGrid) {
|
||||
const wins = [['letzte 1h', d.h1], ['letzte 5h', d.h5],
|
||||
['letzte 24h', d.h24], ['letzte 30 Tage', d.d30]];
|
||||
const z = { calls: 0, tokens_in: 0, tokens_out: 0 };
|
||||
savingsGrid.innerHTML = wins.map(([label, w]) => {
|
||||
const bs = (w && w.by_source) || {};
|
||||
const loc = bs.local || z;
|
||||
const fp = bs['fast-path'] || z;
|
||||
const saved = (loc.calls || 0) + (fp.calls || 0);
|
||||
const locTok = (loc.tokens_in || 0) + (loc.tokens_out || 0);
|
||||
const color = saved > 0 ? '#3FB950' : '#555570';
|
||||
return `<div class="metric-cell">
|
||||
<div class="metric-label">${label}</div>
|
||||
<div class="metric-value" style="color:${color};">${saved} Claude-Calls gespart</div>
|
||||
<div class="metric-sub">lokal ${loc.calls || 0} · fast-path ${fp.calls || 0}</div>
|
||||
<div class="metric-sub">${fmtTokens(locTok)} lokale Tokens (eigene HW)</div>
|
||||
</div>`;
|
||||
}).join('');
|
||||
}
|
||||
|
||||
// 5h-Fenster gegen Plan-Limit: Warn-Klassen
|
||||
const plan = getActivePlanLimit();
|
||||
const limit = plan.h5;
|
||||
@@ -5887,6 +6080,15 @@
|
||||
|
||||
// Vor-definierte Info-Blocks
|
||||
const INFO_TEXTS = {
|
||||
'local-savings': {
|
||||
title: 'Lokales LLM & Claude-Ersparnis',
|
||||
html: `
|
||||
<p>Jeder Turn, den das <strong>lokale LLM</strong> oder ein <strong>Fast-Path</strong> (reiner Skill, ganz ohne LLM) beantwortet, ist ein Claude-Call, der <strong>nicht</strong> gegen dein Subscription-Quota laeuft.</p>
|
||||
<p><strong>„Claude-Calls gespart"</strong> = Anzahl der local- + fast-path-Antworten im Zeitfenster. Konservativ gezaehlt: 1 Antwort = mindestens 1 gesparter Claude-Call (bei Tool-Use waeren es real oft mehr).</p>
|
||||
<p><strong>„lokale Tokens"</strong> = Prompt+Antwort-Tokens, die auf deiner eigenen Gamebox-GPU verarbeitet wurden (echte <code>usage</code>-Zahlen vom Modell, sonst chars/4-Schaetzung). Die kosten dich nichts ausser Strom.</p>
|
||||
<p>Fast-Path-Antworten (z.B. „nächstes Lied") haben ~0 Tokens — reiner Skill-Aufruf, kein Modell.</p>
|
||||
`,
|
||||
},
|
||||
'local-llm': {
|
||||
title: 'Lokales LLM — schnelle Antworten',
|
||||
html: `
|
||||
@@ -6205,18 +6407,41 @@
|
||||
const el = document.getElementById('local-llm-status');
|
||||
if (!el) return;
|
||||
if (!c || !c.enabled) {
|
||||
el.textContent = 'Aus — alle Anfragen laufen ueber Claude.';
|
||||
el.style.color = '#6a6a88';
|
||||
el.textContent = '⚪ Status: AUS — alle Fragen laufen ueber Claude (wie bisher).';
|
||||
el.style.color = '#8888AA';
|
||||
return;
|
||||
}
|
||||
if (c.localOnly) {
|
||||
el.textContent = 'NUR lokal — kein Claude-Fallback (Testmodus). Werkzeug-Turns funktionieren hier nicht.';
|
||||
el.textContent = '🟡 Status: TESTMODUS — nur lokal, Claude ausgesperrt. Werkzeug-Fragen funktionieren nicht.';
|
||||
el.style.color = '#FFD60A';
|
||||
} else {
|
||||
el.textContent = 'Aktiv — einfache Turns lokal (<1s), Rest → Claude.';
|
||||
el.textContent = '🟢 Status: AKTIV — leichte Fragen lokal (<1s), schwere automatisch an Claude.';
|
||||
el.style.color = '#4ADE80';
|
||||
}
|
||||
}
|
||||
let _localModelsCache = [];
|
||||
let _currentLocalModel = 'qwen3-8b';
|
||||
function updateLocalModelDesc() {
|
||||
const el = document.getElementById('local-llm-model-desc');
|
||||
const sel = document.getElementById('local-llm-model');
|
||||
if (!el || !sel) return;
|
||||
const m = _localModelsCache.find(x => x.id === sel.value);
|
||||
el.textContent = m && m.description ? m.description : '';
|
||||
}
|
||||
async function loadLocalModelList() {
|
||||
try {
|
||||
const r = await fetch('/api/local-models-list');
|
||||
const j = await r.json();
|
||||
_localModelsCache = (j && j.models) || [];
|
||||
} catch (e) { _localModelsCache = []; }
|
||||
const sel = document.getElementById('local-llm-model');
|
||||
if (sel) {
|
||||
sel.innerHTML = _localModelsCache.map(m =>
|
||||
`<option value="${m.id}">${m.display_name || m.id}</option>`).join('') || '<option value="">(keine)</option>';
|
||||
if (_localModelsCache.some(m => m.id === _currentLocalModel)) sel.value = _currentLocalModel;
|
||||
updateLocalModelDesc();
|
||||
}
|
||||
}
|
||||
async function loadLocalLlmConfig() {
|
||||
try {
|
||||
const r = await fetch('/api/local-llm-config');
|
||||
@@ -6227,15 +6452,24 @@
|
||||
if (en) en.checked = !!c.enabled;
|
||||
if (ol) ol.checked = !!c.localOnly;
|
||||
if (tv) tv.value = (c.toolVariant === 'full') ? 'full' : 'slim';
|
||||
_currentLocalModel = c.localLlmModel || 'qwen3-8b';
|
||||
const sel = document.getElementById('local-llm-model');
|
||||
if (sel && _localModelsCache.some(m => m.id === _currentLocalModel)) {
|
||||
sel.value = _currentLocalModel;
|
||||
updateLocalModelDesc();
|
||||
}
|
||||
setLocalLlmStatus(c);
|
||||
} catch (e) { /* still */ }
|
||||
}
|
||||
async function saveLocalLlmConfig() {
|
||||
const modelSel = document.getElementById('local-llm-model');
|
||||
const body = {
|
||||
enabled: document.getElementById('local-llm-enabled').checked,
|
||||
localOnly: document.getElementById('local-llm-onlylocal').checked,
|
||||
toolVariant: document.getElementById('local-llm-toolvariant').value,
|
||||
localLlmModel: (modelSel && modelSel.value) || '',
|
||||
};
|
||||
updateLocalModelDesc();
|
||||
try {
|
||||
const r = await fetch('/api/local-llm-config', {
|
||||
method: 'POST', headers: { 'Content-Type': 'application/json' },
|
||||
|
||||
+62
-3
@@ -309,9 +309,10 @@ function readLocalLlmConfig() {
|
||||
enabled: !!p.enabled,
|
||||
localOnly: !!p.localOnly,
|
||||
toolVariant: p.toolVariant === "full" ? "full" : "slim",
|
||||
localLlmModel: (typeof p.localLlmModel === "string" && p.localLlmModel) ? p.localLlmModel : "qwen3-8b",
|
||||
};
|
||||
} catch {
|
||||
return { enabled: false, localOnly: false, toolVariant: "slim" };
|
||||
return { enabled: false, localOnly: false, toolVariant: "slim", localLlmModel: "qwen3-8b" };
|
||||
}
|
||||
}
|
||||
function writeLocalLlmConfig(patch) {
|
||||
@@ -319,6 +320,7 @@ function writeLocalLlmConfig(patch) {
|
||||
if (typeof patch.enabled === "boolean") cur.enabled = patch.enabled;
|
||||
if (typeof patch.localOnly === "boolean") cur.localOnly = patch.localOnly;
|
||||
if (patch.toolVariant === "slim" || patch.toolVariant === "full") cur.toolVariant = patch.toolVariant;
|
||||
if (typeof patch.localLlmModel === "string" && patch.localLlmModel.trim()) cur.localLlmModel = patch.localLlmModel.trim();
|
||||
fs.mkdirSync("/shared/config", { recursive: true });
|
||||
const tmp = LOCAL_LLM_CONFIG_FILE + ".tmp";
|
||||
fs.writeFileSync(tmp, JSON.stringify(cur, null, 2));
|
||||
@@ -326,6 +328,27 @@ function writeLocalLlmConfig(patch) {
|
||||
return cur;
|
||||
}
|
||||
|
||||
// ── Lokale Modell-Liste (Diagnostic-Dropdown) ────────────────
|
||||
// /shared/config/local_models.json — kuratierte Liste; muss zu den KEYS in
|
||||
// xtts/llama-swap/config.yaml passen. Wird bei Bedarf mit Defaults seeded.
|
||||
const LOCAL_MODELS_FILE = "/shared/config/local_models.json";
|
||||
const DEFAULT_LOCAL_MODELS = [
|
||||
{ id: "qwen3-8b", display_name: "Qwen3 8B (Standard)", description: "Bestes Tool-Calling, ~6 GB. Passt auf 12 GB." },
|
||||
{ id: "qwen3-4b", display_name: "Qwen3 4B (schneller)", description: "Kleiner + flotter, ~3 GB. Etwas schwaecher." },
|
||||
];
|
||||
function loadLocalModels() {
|
||||
try {
|
||||
const arr = JSON.parse(fs.readFileSync(LOCAL_MODELS_FILE, "utf-8"));
|
||||
if (Array.isArray(arr) && arr.length && arr.every(m => m && typeof m.id === "string")) return arr;
|
||||
} catch {}
|
||||
// Seed defaults
|
||||
try {
|
||||
fs.mkdirSync("/shared/config", { recursive: true });
|
||||
fs.writeFileSync(LOCAL_MODELS_FILE, JSON.stringify(DEFAULT_LOCAL_MODELS, null, 2));
|
||||
} catch {}
|
||||
return DEFAULT_LOCAL_MODELS;
|
||||
}
|
||||
|
||||
// ── File-Project-Manifest ───────────────────────────────────────────
|
||||
// Jeder Eintrag map[absoluter_pfad] = project_id (leer = Hauptchat).
|
||||
// Wird vom files-list-Endpoint + files-set-project gepflegt.
|
||||
@@ -875,6 +898,11 @@ function connectRVS(forcePlain) {
|
||||
// Mode-Broadcast von der Bridge → an Browser-Clients weiterreichen
|
||||
log("info", "rvs", `Mode-Broadcast: ${msg.payload?.mode} (${msg.payload?.name})`);
|
||||
broadcast({ type: "mode", payload: msg.payload });
|
||||
} else if (msg.type === "project_changed") {
|
||||
// Ein Projekt wurde geaendert (ARIA-Tool, App-Verstecken, …) → an die
|
||||
// Browser-Tabs weiterreichen, damit die Projektliste live neu laedt
|
||||
// (bisher wurde das NICHT geforwardet → Diagnostic aktualisierte nie).
|
||||
broadcast({ type: "project_changed", payload: msg.payload || {} });
|
||||
} else if (msg.type === "agent_activity") {
|
||||
// Bridge meldet "ARIA denkt/schreibt/tool" oder "idle" — an Browser
|
||||
// weiterreichen, damit der Thinking-Indikator im Chat erscheint.
|
||||
@@ -1574,7 +1602,16 @@ const htmlPath = path.join(__dirname, "index.html");
|
||||
|
||||
const server = http.createServer((req, res) => {
|
||||
if (req.url === "/" || req.url === "/index.html") {
|
||||
res.writeHead(200, { "Content-Type": "text/html; charset=utf-8" });
|
||||
// no-store: das Dashboard ist eine Single-HTML-App die bei jedem Deploy
|
||||
// neue Inline-JS/CSS bekommt. Ohne Cache-Header servierte der Browser die
|
||||
// alte Version trotz Reload (neue Features tauchten erst nach Hard-Reload
|
||||
// auf) — genau das Symptom „ich seh den Button nicht".
|
||||
res.writeHead(200, {
|
||||
"Content-Type": "text/html; charset=utf-8",
|
||||
"Cache-Control": "no-store, no-cache, must-revalidate",
|
||||
"Pragma": "no-cache",
|
||||
"Expires": "0",
|
||||
});
|
||||
res.end(fs.readFileSync(htmlPath, "utf-8"));
|
||||
} else if (req.url === "/api/state") {
|
||||
res.writeHead(200, { "Content-Type": "application/json" });
|
||||
@@ -1602,6 +1639,9 @@ const server = http.createServer((req, res) => {
|
||||
}
|
||||
});
|
||||
return;
|
||||
} else if (req.url === "/api/local-models-list" && req.method === "GET") {
|
||||
res.writeHead(200, { "Content-Type": "application/json" });
|
||||
res.end(JSON.stringify({ ok: true, models: loadLocalModels() }));
|
||||
} else if (req.url === "/api/local-llm-config" && req.method === "GET") {
|
||||
res.writeHead(200, { "Content-Type": "application/json" });
|
||||
res.end(JSON.stringify(readLocalLlmConfig()));
|
||||
@@ -2142,6 +2182,13 @@ const server = http.createServer((req, res) => {
|
||||
// mehr als eine Minute.
|
||||
const isUpload = /\/attachments(\/upload)?$/.test(targetPath);
|
||||
const timeout = isUpload ? 120000 : 60000;
|
||||
// Projekt-Mutationen (create/switch/end/archive/patch inkl. hidden) sollen
|
||||
// alle Clients live aktualisieren. Wir broadcasten nach Erfolg ein
|
||||
// project_changed an RVS — App + andere Diagnostic-Tabs laden dann neu,
|
||||
// ohne Seiten-Refresh (spiegelt das bestehende ARIA-project_changed-Event).
|
||||
const isProjectMutation =
|
||||
/^\/projects\b/.test(targetPath) &&
|
||||
(req.method === "POST" || req.method === "PATCH" || req.method === "DELETE");
|
||||
const proxyReq = http.request({
|
||||
host: "aria-brain",
|
||||
port: 8080,
|
||||
@@ -2152,6 +2199,17 @@ const server = http.createServer((req, res) => {
|
||||
}, (proxyRes) => {
|
||||
res.writeHead(proxyRes.statusCode, proxyRes.headers);
|
||||
proxyRes.pipe(res);
|
||||
if (isProjectMutation && proxyRes.statusCode >= 200 && proxyRes.statusCode < 300) {
|
||||
try {
|
||||
// An App + Bridge (RVS echot NICHT an den Sender) …
|
||||
sendToRVS_raw({ type: "project_changed",
|
||||
payload: { reason: "diagnostic" },
|
||||
timestamp: Date.now() });
|
||||
// … und an die eigenen Browser-Tabs (die haengen am Diag-Server, nicht
|
||||
// direkt am RVS, kriegen den RVS-Broadcast also nicht).
|
||||
broadcast({ type: "project_changed", payload: { reason: "diagnostic" } });
|
||||
} catch (_) {}
|
||||
}
|
||||
});
|
||||
proxyReq.on("error", (err) => {
|
||||
res.writeHead(503, { "Content-Type": "application/json" });
|
||||
@@ -2816,6 +2874,7 @@ async function handleLoadChatHistory(clientWs) {
|
||||
const ts = obj.ts || 0;
|
||||
const text = String(obj.text || "");
|
||||
const projectId = String(obj.project_id || ""); // Multi-Threading: Kontext-Zuordnung
|
||||
const answeredBy = String(obj.answeredBy || ""); // Quell-Badge (local/claude/fast-path)
|
||||
if (obj.role === "user") {
|
||||
if (text) messages.push({ type: "sent", text, meta: "Gateway direkt", ts, projectId });
|
||||
continue;
|
||||
@@ -2842,7 +2901,7 @@ async function handleLoadChatHistory(clientWs) {
|
||||
projectId,
|
||||
});
|
||||
}
|
||||
if (text) messages.push({ type: "received", text, meta: "chat:final", ts, projectId });
|
||||
if (text) messages.push({ type: "received", text, meta: "chat:final", ts, projectId, answeredBy });
|
||||
}
|
||||
|
||||
clientWs.send(JSON.stringify({ type: "chat_history", messages }));
|
||||
|
||||
@@ -53,6 +53,21 @@ services:
|
||||
networks:
|
||||
- aria-net
|
||||
|
||||
# ─── SearXNG (self-hosted Meta-Suche) ────────────────────
|
||||
# Backend fuer das web_search-Tool (B1b). Aggregiert Google/Bing/Brave/… ohne
|
||||
# API-Key, laeuft nur intern auf aria-net. Config: aria-data/searxng/settings.yml
|
||||
# (JSON-Format aktiviert, Rate-Limiter aus fuer den Brain-Zugriff).
|
||||
searxng:
|
||||
image: searxng/searxng:latest
|
||||
container_name: aria-searxng
|
||||
volumes:
|
||||
- ./aria-data/searxng:/etc/searxng
|
||||
environment:
|
||||
- SEARXNG_BASE_URL=http://searxng:8080/
|
||||
restart: unless-stopped
|
||||
networks:
|
||||
- aria-net
|
||||
|
||||
# ─── ARIA Brain (Agent + Memory) ─────────────────────────
|
||||
# Loest das alte aria-core (OpenClaw) ab. Vector-DB-basiertes
|
||||
# Memory, eigener Agent-Loop, SSH zur aria-wohnung-VM.
|
||||
@@ -86,6 +101,8 @@ services:
|
||||
- RVS_HOST=${RVS_HOST:-}
|
||||
- RVS_PORT_PUBLIC=${RVS_PORT_PUBLIC:-${RVS_PORT:-443}}
|
||||
- RVS_TLS=${RVS_TLS:-true}
|
||||
# SearXNG (self-hosted Meta-Suche) fuer das web_search-Tool (B1b).
|
||||
- SEARXNG_URL=${SEARXNG_URL:-http://searxng:8080}
|
||||
volumes:
|
||||
- ./aria-data/brain/data:/data # Memory-Cache + Skills + Models (bind-mount fuer Export)
|
||||
- ./aria-data/brain-import:/import:ro # Quell-MDs fuer den initialen Memory-Import (read-only)
|
||||
|
||||
+20
-24
@@ -90,22 +90,19 @@ services:
|
||||
# Container-Restarts.
|
||||
restart: unless-stopped
|
||||
|
||||
# ─── Lokales LLM (Plan B, B0) — llama.cpp-Server (GPU) ────────
|
||||
# Serviert Qwen3-8B (GGUF Q4_K_M) OpenAI-kompatibel auf :8081, NUR im
|
||||
# Compose-Netz (kein RVS direkt) — die Bruecke macht der llm-adapter.
|
||||
# ─── Lokales LLM (Plan B, B0.5) — llama-swap (GPU) ────────────
|
||||
# llama-swap laedt/swappt mehrere Modelle on-demand (nur eins passt gleich-
|
||||
# zeitig in die 12 GB). Welches geladen wird, bestimmt das `model`-Feld im
|
||||
# Request — das Brain schickt es aus local_llm.json mit. Erster Load eines
|
||||
# Modells zieht das GGUF via -hf von HF (Cache unter /models, persistent).
|
||||
# OpenAI-kompatibel auf :8080, nur im Compose-Netz; die Bruecke macht der
|
||||
# llm-adapter. Modell-Liste: ./llama-swap/config.yaml.
|
||||
#
|
||||
# AUTO-DOWNLOAD: llama.cpp zieht das GGUF beim ersten Start selbst von
|
||||
# Hugging Face (-hf <repo>:<quant>) und cached es unter /models (persistent
|
||||
# via Bind-Mount -> kein Re-Download bei Restart). Kein manuelles Ablegen
|
||||
# noetig. Modell wechseln = LLM_HF_REPO/LLM_HF_QUANT in der .env aendern +
|
||||
# Container neu. (Alternativ lokale Datei: command auf -m /models/x.gguf.)
|
||||
#
|
||||
# VRAM auf der RTX 3060 (12 GB): whisper-small (~1-2) + f5tts (~1-2) +
|
||||
# qwen3-8b-q4 (~6) ~= 9-10 GB. Passt, aber knapp — bei OOM: LLM_CTX kleiner
|
||||
# oder Quant auf Q4_K_S/IQ4_XS wechseln.
|
||||
llama:
|
||||
image: ghcr.io/ggml-org/llama.cpp:server-cuda
|
||||
container_name: aria-llama
|
||||
# BLIND GEBAUT (kein Gamebox-Test hier): beim ersten Start
|
||||
# `docker logs -f aria-llama-swap` pruefen. Image bundelt llama-server.
|
||||
llama-swap:
|
||||
image: ghcr.io/mostlygeek/llama-swap:unified-cuda
|
||||
container_name: aria-llama-swap
|
||||
deploy:
|
||||
resources:
|
||||
reservations:
|
||||
@@ -114,13 +111,11 @@ services:
|
||||
count: 1
|
||||
capabilities: [gpu]
|
||||
volumes:
|
||||
- ./models:/models # HF-Download-Cache (persistent)
|
||||
- ./models:/models # HF-Download-Cache (persistent)
|
||||
- ./llama-swap/config.yaml:/app/config.yaml:ro # Modell-Liste
|
||||
environment:
|
||||
- LLAMA_CACHE=/models # llama.cpp legt -hf-Downloads hier ab
|
||||
command: >
|
||||
-hf ${LLM_HF_REPO:-Qwen/Qwen3-8B-GGUF}:${LLM_HF_QUANT:-Q4_K_M}
|
||||
--host 0.0.0.0 --port 8081
|
||||
-ngl 99 -c ${LLM_CTX:-8192} --jinja
|
||||
- LLAMA_CACHE=/models # llama-server legt -hf-Downloads hier ab
|
||||
command: ["--config", "/app/config.yaml", "--listen", "0.0.0.0:8080"]
|
||||
restart: unless-stopped
|
||||
|
||||
# ─── Local-LLM-Adapter — RVS <-> llama.cpp (Plan B, B0) ──────
|
||||
@@ -130,14 +125,15 @@ services:
|
||||
build: ./llm-adapter
|
||||
container_name: aria-llm-adapter
|
||||
depends_on:
|
||||
- llama
|
||||
- llama-swap
|
||||
environment:
|
||||
- RVS_HOST=${RVS_HOST}
|
||||
- RVS_PORT=${RVS_PORT:-443}
|
||||
- RVS_TLS=${RVS_TLS:-true}
|
||||
- RVS_TLS_FALLBACK=${RVS_TLS_FALLBACK:-true}
|
||||
- RVS_TOKEN=${RVS_TOKEN}
|
||||
- LLAMA_URL=http://llama:8081
|
||||
- LLAMA_URL=http://llama-swap:8080
|
||||
- LLM_MODEL=${LLM_MODEL:-qwen3-8b}
|
||||
- LLM_TIMEOUT_SEC=${LLM_TIMEOUT_SEC:-60}
|
||||
# Erster Load eines Modells kann ein GGUF ziehen (mehrere GB) — grosszuegig.
|
||||
- LLM_TIMEOUT_SEC=${LLM_TIMEOUT_SEC:-600}
|
||||
restart: unless-stopped
|
||||
|
||||
@@ -0,0 +1,42 @@
|
||||
# llama-swap Modell-Liste fuer ARIA (Plan B, B0.5).
|
||||
# Welches Modell geladen wird, bestimmt das `model`-Feld im Request (das Brain
|
||||
# schickt es aus /shared/config/local_llm.json mit). llama-swap laedt es
|
||||
# on-demand, swappt bei Bedarf (nur eins passt gleichzeitig in die 12 GB).
|
||||
# Erster Load zieht das GGUF via -hf von Hugging Face (Cache unter /models).
|
||||
#
|
||||
# Die Modell-KEYS hier muessen zu local_models.json (Diagnostic-Dropdown) passen.
|
||||
#
|
||||
# healthCheckTimeout: Sekunden, die llama-swap auf "Modell bereit" wartet.
|
||||
# GROSSZUEGIG, weil der erste Load ein GGUF (mehrere GB) herunterlaedt. Wenn der
|
||||
# erste Download laenger dauert und abbricht: hier hochsetzen.
|
||||
healthCheckTimeout: 1800
|
||||
|
||||
models:
|
||||
# Standard — Qwen3 8B (~6 GB Q4). Bestes Tool-Calling, passt auf 12 GB.
|
||||
"qwen3-8b":
|
||||
cmd: |
|
||||
llama-server --port ${PORT} --host 127.0.0.1
|
||||
-hf Qwen/Qwen3-8B-GGUF:Q4_K_M
|
||||
-ngl 99 -c 8192 --jinja
|
||||
ttl: 3600 # nach 1h Idle entladen (VRAM freigeben)
|
||||
|
||||
# Kleiner + schneller — Qwen3 4B (~3 GB). Fuer noch flottere Antworten,
|
||||
# etwas schwaecher. Guter A/B-Vergleich gegen 8B.
|
||||
"qwen3-4b":
|
||||
cmd: |
|
||||
llama-server --port ${PORT} --host 127.0.0.1
|
||||
-hf Qwen/Qwen3-4B-GGUF:Q4_K_M
|
||||
-ngl 99 -c 8192 --jinja
|
||||
ttl: 3600
|
||||
|
||||
# ── Vorlagen fuer spaeter (auskommentiert; brauchen mehr VRAM / 2. Karte) ──
|
||||
# "qwen3-14b":
|
||||
# cmd: |
|
||||
# llama-server --port ${PORT} --host 127.0.0.1
|
||||
# -hf Qwen/Qwen3-14B-GGUF:Q4_K_M -ngl 99 -c 8192 --jinja
|
||||
# ttl: 3600
|
||||
# "mistral-small-3":
|
||||
# cmd: |
|
||||
# llama-server --port ${PORT} --host 127.0.0.1
|
||||
# -hf <mistral-small-3-gguf-repo>:Q4_K_M -ngl 99 -c 8192 --jinja
|
||||
# ttl: 3600
|
||||
+63
-11
@@ -69,11 +69,16 @@ async def _send(ws, mtype: str, payload: dict) -> None:
|
||||
|
||||
|
||||
async def _call_llama(messages: list, *, max_tokens: int, temperature: float,
|
||||
stop) -> dict:
|
||||
"""Ruft llama.cpp /v1/chat/completions (OpenAI-Format). Gibt
|
||||
{ok, content, error} zurueck — wirft nie."""
|
||||
stop, tools=None, model=None) -> dict:
|
||||
"""Ruft llama.cpp/llama-swap /v1/chat/completions (OpenAI-Format). Gibt
|
||||
{ok, content, tool_calls, error} zurueck — wirft nie.
|
||||
|
||||
model: welches Modell llama-swap laden soll (B0.5). Kommt aus dem Request
|
||||
(Brain -> local_llm.json). Faellt auf LLM_MODEL (env) zurueck.
|
||||
tools: optionale OpenAI-Tool-Definitionen (B1b). Qwen3 (--jinja) kann
|
||||
natives Tool-Calling und liefert dann message.tool_calls."""
|
||||
body = {
|
||||
"model": LLM_MODEL,
|
||||
"model": model or LLM_MODEL,
|
||||
"messages": messages,
|
||||
"max_tokens": max_tokens,
|
||||
"temperature": temperature,
|
||||
@@ -81,6 +86,9 @@ async def _call_llama(messages: list, *, max_tokens: int, temperature: float,
|
||||
}
|
||||
if stop:
|
||||
body["stop"] = stop
|
||||
if tools:
|
||||
body["tools"] = tools
|
||||
body["tool_choice"] = "auto"
|
||||
if LLM_DISABLE_THINKING:
|
||||
# llama.cpp (--jinja) reicht chat_template_kwargs an die Chat-Vorlage
|
||||
# weiter. Qwen3 unterdrueckt damit den <think>-Block.
|
||||
@@ -90,14 +98,34 @@ async def _call_llama(messages: list, *, max_tokens: int, temperature: float,
|
||||
r = await client.post(f"{LLAMA_URL}/v1/chat/completions", json=body)
|
||||
r.raise_for_status()
|
||||
data = r.json()
|
||||
content = (data.get("choices") or [{}])[0].get("message", {}).get("content", "")
|
||||
return {"ok": True, "content": content or "", "usage": data.get("usage")}
|
||||
msg = (data.get("choices") or [{}])[0].get("message", {}) or {}
|
||||
return {
|
||||
"ok": True,
|
||||
"content": msg.get("content") or "",
|
||||
"tool_calls": msg.get("tool_calls") or None,
|
||||
"usage": data.get("usage"),
|
||||
}
|
||||
except Exception as e:
|
||||
logger.warning("llama.cpp-Call fehlgeschlagen: %s", e)
|
||||
return {"ok": False, "content": "", "error": str(e)[:300]}
|
||||
|
||||
|
||||
# B0.5-2: Lade-Status ans Diagnostic (service_status, service="llm"). Wir kennen
|
||||
# den Download-Fortschritt nicht (llama-swap gibt ihn nicht her), aber wir melden
|
||||
# den Zustand bei Modellwechsel: loading -> ready/error. _last_model = aktuell
|
||||
# geladenes; _ready_models = in dieser Session schon einmal bereit gewesene
|
||||
# (fuer den "frisch geladen"-Hinweis 🎉 bei langem Erst-Load).
|
||||
_last_model = None
|
||||
_ready_models: set = set()
|
||||
|
||||
|
||||
async def _emit_llm_status(ws, state: str, model: str, **extra) -> None:
|
||||
await _send(ws, "service_status",
|
||||
{"service": "llm", "state": state, "model": model, **extra})
|
||||
|
||||
|
||||
async def _handle_llm_request(ws, payload: dict) -> None:
|
||||
global _last_model
|
||||
req_id = payload.get("requestId", "")
|
||||
messages = payload.get("messages") or []
|
||||
if not isinstance(messages, list) or not messages:
|
||||
@@ -108,19 +136,43 @@ async def _handle_llm_request(ws, payload: dict) -> None:
|
||||
max_tokens = int(payload.get("max_tokens", 512) or 512)
|
||||
temperature = float(payload.get("temperature", 0.7) or 0.7)
|
||||
stop = payload.get("stop")
|
||||
tools = payload.get("tools") or None
|
||||
model = (payload.get("model") or "").strip() or None
|
||||
eff_model = model or LLM_MODEL
|
||||
|
||||
# Modellwechsel (oder erster Request) → llama-swap laedt/swappt: Status melden.
|
||||
switching = eff_model != _last_model
|
||||
if switching:
|
||||
await _emit_llm_status(ws, "loading", eff_model)
|
||||
|
||||
t0 = time.time()
|
||||
res = await _call_llama(messages, max_tokens=max_tokens,
|
||||
temperature=temperature, stop=stop)
|
||||
temperature=temperature, stop=stop, tools=tools,
|
||||
model=model)
|
||||
dt = time.time() - t0
|
||||
logger.info("llm_request id=%s -> ok=%s %.2fs content_len=%d",
|
||||
(req_id[:8] if req_id else "?"), res.get("ok"), dt,
|
||||
len(res.get("content") or ""))
|
||||
|
||||
if switching:
|
||||
if res.get("ok"):
|
||||
fresh = (eff_model not in _ready_models) and dt > 25
|
||||
_ready_models.add(eff_model)
|
||||
_last_model = eff_model
|
||||
await _emit_llm_status(ws, "ready", eff_model,
|
||||
loadSeconds=round(dt, 1), freshlyDownloaded=fresh)
|
||||
else:
|
||||
# bei Fehler _last_model NICHT setzen → naechster Versuch meldet erneut loading
|
||||
await _emit_llm_status(ws, "error", eff_model,
|
||||
error=(res.get("error") or "")[:120])
|
||||
tc = res.get("tool_calls")
|
||||
logger.info("llm_request id=%s model=%s -> ok=%s %.2fs content_len=%d tool_calls=%d",
|
||||
(req_id[:8] if req_id else "?"), model or LLM_MODEL, res.get("ok"), dt,
|
||||
len(res.get("content") or ""), len(tc) if tc else 0)
|
||||
await _send(ws, "llm_response", {
|
||||
"requestId": req_id,
|
||||
"ok": res.get("ok", False),
|
||||
"content": res.get("content", ""),
|
||||
"tool_calls": tc,
|
||||
"error": res.get("error"),
|
||||
"model": LLM_MODEL,
|
||||
"model": model or LLM_MODEL,
|
||||
"elapsedMs": int(dt * 1000),
|
||||
})
|
||||
|
||||
|
||||
Reference in New Issue
Block a user