v0.2: in-weights option (Meta Muse Glimmer 30B + ViKernel, mode A/B)

ag.js

@@ -5,6 +5,7 @@
 'use strict';
 const S = {
   en: {
+    agGraftCold: 'The in-weights backend (GPU) is starting. Try again in a few minutes.', agEGate: 'E checks the audit chain of the ViKernel gate.', agModeA: 'Mode A: the ViKernel kernel runs inside the model weights (a chip circuit grafted into the decoder). Each turn takes about 20 to 60 seconds.',
     agName: 'Agent Guard (Meta Muse)',
     agIntro: 'A Meta Muse agent. Every tool call it proposes is sent to the ViKernel Secure OS first. Denied or halted actions never run.',
     agChat: 'Agent chat', agDec: 'ViKernel decisions', agAudit: 'Audit chain',
@@ -22,6 +23,7 @@ const S = {
     agStats: 'Live counters', agMedGate: 'median gate', agMedModel: 'median model turn', agTok: 'tokens', agFoot: 'All tools are mocks with fake data. The kernel runs as a private service on VIDRAFT infrastructure.',
   },
   ko: {
+    agGraftCold: '가중치 내장 백엔드(GPU)가 시작 중입니다. 몇 분 뒤 다시 시도하세요.', agEGate: 'E 는 ViKernel 게이트의 감사 사슬을 검사합니다.', agModeA: '모드 A: ViKernel 커널이 모델 가중치 안(디코더에 접목한 칩 회로)에서 동작합니다. 한 번에 약 20~60초 걸립니다.',
     agName: 'Agent Guard (Meta Muse)',
     agIntro: 'Meta Muse 에이전트입니다. 에이전트가 제안하는 모든 도구 호출은 먼저 ViKernel 시큐어 OS 로 갑니다. 거부되거나 정지된 행동은 실행되지 않습니다.',
     agChat: '에이전트 대화', agDec: 'ViKernel 판정', agAudit: '감사 사슬',
@@ -46,12 +48,13 @@ const ICON = '<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-
 window.VK_AG = ({el, t, toast}) => ({
   id: 'guard', name: () => t('agName'), color: '#0866FF', icon: ICON, w: 900, h: 560,
   render(b) {
-    let LIVE = false, REPLAY = null, busy = false;
+    let LIVE = false, REPLAY = null, busy = false, SEL = 'spark';
+    const GRAFT = () => SEL !== 'spark', MODE = () => SEL === 'graftA' ? 'A' : 'B';
     const SID = 's' + Math.random().toString(36).slice(2, 8);
-    const call = async (path, body) => {
-      const ctl = new AbortController(); const to = setTimeout(() => ctl.abort(), 120000);
+    const call = async (path, body, spark) => {
+      const ctl = new AbortController(); const to = setTimeout(() => ctl.abort(), 320000);
       try {
-        const r = await fetch('api/ag/' + path, body === undefined ? {cache:'no-store', signal:ctl.signal} :
+        const r = await fetch((GRAFT() && !spark ? 'api/agx/' : 'api/ag/') + path, body === undefined ? {cache:'no-store', signal:ctl.signal} :
           {method:'POST', headers:{'Content-Type':'application/json'}, body:JSON.stringify(body), signal:ctl.signal});
         if (r.status === 429) { toast(t('agRate')); throw new Error('rate'); }
         if (!r.ok) throw new Error('http ' + r.status);
@@ -60,6 +63,16 @@ window.VK_AG = ({el, t, toast}) => ({
     };
     const mode = el('span', {class:'tag rec'}, '...');
     const stateTag = el('span', {class:'tag none'}, 'kernel: ...');
+    const sel = el('select', {class:'aginput', style:'flex:0 1 auto;height:28px;max-width:100%', 'aria-label':'model'},
+      el('option', {value:'spark'}, '1. Meta Muse Spark 1.3 (API) + ViKernel gate'),
+      el('option', {value:'graftB'}, '2. Meta Muse Glimmer 30B + ViKernel in-weights, mode B'),
+      el('option', {value:'graftA'}, '3. Meta Muse Glimmer 30B + ViKernel in-weights, mode A (kernel runs inside the weights, slow)'));
+    sel.onchange = async () => { const want = sel.value;
+      if (want !== 'spark') { const prev = SEL; SEL = want; let h = null; try { h = await call('health'); } catch {}
+        if (!h || !h.live) { SEL = prev; sel.value = prev; toast(t('agGraftCold')); return; } }
+      SEL = want; LIVE = true; REPLAY = null; mode.textContent = t('agLive'); mode.className = 'tag allow';
+      b.querySelector('#agModel').textContent = sel.options[sel.selectedIndex].text.replace(/^\d\. /, '');
+      say('s', '--- ' + sel.options[sel.selectedIndex].text + ' ---' + (want === 'graftA' ? '\n' + t('agModeA') : '')); audit.innerHTML = ''; refresh(); };
     const chat = el('div', {class:'agchat'}), stream = el('div', {class:'agstream'}), audit = el('div', {class:'agaudit mono small'});
     const msg = el('input', {class:'aginput', placeholder:t('agPh'), autocomplete:'off'});
     const key = el('input', {class:'aginput mono', type:'password', placeholder:t('agKey'), style:'flex:1 1 90px;max-width:160px;min-width:80px'});
@@ -92,6 +105,7 @@ window.VK_AG = ({el, t, toast}) => ({
       const lines = Object.entries(r).filter(([k]) => !['pass', 'note', 'first_error', 'seconds'].includes(k)).map(([k, v]) => k + ': ' + (typeof v === 'object' ? JSON.stringify(v) : v));
       say('s', [el('b', {style:`color:${r.pass ? '#065F46' : 'var(--rose)'}`}, (r.pass ? 'PASS ' : 'FAIL ') + name), '\n' + lines.join('\n')]); };
     const refresh = async () => { if (!LIVE) return;
+      if (GRAFT()) { try { const s = await call('status'); stateTag.textContent = 'kernel (in weights): ' + s.kernel_state; stateTag.className = 'tag ' + (s.kernel_state === 'RUN' ? 'allow' : 'deny'); } catch {} return; }
       try { const s = await call('status'); stateTag.textContent = 'kernel: ' + s.kernel_state; stateTag.className = 'tag ' + (s.kernel_state === 'RUN' ? 'allow' : 'deny');
         const a = await call('audit'); const v = a.verify; audit.innerHTML = '';
         audit.append(el('div', {style:`font-weight:700;color:${v.ok ? '#065F46' : 'var(--rose)'}`}, (v.ok ? t('agOk') : t('agBroken')) + ' (' + v.n + ' ' + t('agRecords') + ')'),
@@ -104,19 +118,19 @@ window.VK_AG = ({el, t, toast}) => ({
     const lock = on => { busy = on; b.querySelectorAll('button').forEach(x => x.disabled = on); };
     const scen = async name => { if (busy) return; lock(true); say('s', '--- ' + name + ' ---');
       try {
-        if (LIVE) { try { const r = await call('scenario', {name}); show(r.events); result(name, r.result); } catch (e) { if (e.message !== 'rate') { await goReplay(); } } }
+        if (LIVE) { try { const ex = GRAFT() && name === 'e_tamper'; if (ex) say('s', t('agEGate')); const r = await call('scenario', {name, mode:MODE()}, ex); show(r.events); result(name, r.result); } catch (e) { if (e.message !== 'rate') { await goReplay(); } } }
         if (!LIVE) { const s = REPLAY.scenarios[name]; if (s) { for (const e of s.events) { show([e]); await new Promise(r => setTimeout(r, 300)); } result(name, s.result); } }
       } finally { lock(false); refresh(); } };
     const send = async () => { const m = msg.value.trim(); if (!m || busy) return; msg.value = '';
       if (!LIVE) { say('s', t('agReplayChat')); return; }
       lock(true); say('s', t('agThinking'));
-      try { const r = await call('chat', {message:m, session:SID}); show(r.events); } catch (e) { if (e.message !== 'rate') say('s', t('agCold')); }
+      try { const r = await call('chat', {message:m, session:SID, mode:MODE()}); show(r.events); } catch (e) { if (e.message !== 'rate') say('s', t('agCold')); }
       finally { lock(false); refresh(); } };
     msg.addEventListener('keydown', e => { if (e.key === 'Enter' && !e.isComposing) send(); });
     const btn = (k, f, cls) => el('button', {class:'btn ' + (cls || ''), onclick:f}, t(k));
     b.append(el('div', {class:'aggrid'},
       el('div', {class:'agcol'},
-        el('div', {class:'row', style:'margin-bottom:6px;flex-wrap:wrap;gap:6px'}, el('b', {}, t('agChat')), mode, el('span', {class:'tag none', id:'agModel'}, 'Meta Muse Spark 1.3')),
+        el('div', {class:'row', style:'margin-bottom:6px;flex-wrap:wrap;gap:6px'}, el('b', {}, t('agChat')), mode, el('span', {class:'tag none', id:'agModel'}, 'Meta Muse Spark 1.3'), sel),
         el('div', {class:'row', style:'flex-wrap:wrap;gap:6px;margin-bottom:6px'}, ['a_normal', 'b_injection', 'c_persuasion', 'd_kill', 'e_tamper'].map((s, i) => btn('ag' + 'ABCDE'[i], () => scen(s)))),
         el('p', {class:'small muted', style:'margin:0 0 6px'}, t('agIntro')),
         el('details', {class:'card small', style:'margin:0 0 6px;padding:6px 10px', open:''}, el('summary', {style:'font-weight:700;cursor:pointer'}, t('agTry')),

server.py

@@ -1,94 +1,100 @@
-"""ViKernel Desktop Space server: serves the static desktop and proxies /api/ag/* server-side
-to a private Agent Guard backend. Backend credentials live in Space secrets and never reach the browser."""
-import json, os, time, urllib.error, urllib.request
-from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer
-
-BACKEND = os.environ.get("BACKEND_URL", "").rstrip("/")
-HF_AUTH = os.environ.get("BACKEND_HF_TOKEN", "")
-BK_TOKEN = os.environ.get("BACKEND_TOKEN", "")
-RATE = int(os.environ.get("RATE_PER_10MIN", "20"))
-TIMEOUT = float(os.environ.get("BACKEND_TIMEOUT", "110"))
-GET_OK = {"health", "status", "audit"}
-POST_OK = {"chat", "scenario", "kill", "resume", "tamper", "reset"}
-STATIC = {"index.html", "app.js", "ag.js", "i18n.js", "demo_data.js", "styles.css", "ag_replay.json"}
-HITS = {}
-
-
-def limited(ip):
-    now = time.time()
-    h = [x for x in HITS.get(ip, []) if now - x < 600]
-    full = len(h) >= RATE
-    if not full:
-        h.append(now)
-    HITS[ip] = h
-    if len(HITS) > 5000:
-        HITS.clear()
-    return full
-
-
-class H(SimpleHTTPRequestHandler):
-    def _json(self, code, obj):
-        b = json.dumps(obj).encode()
-        self.send_response(code); self.send_header("Content-Type", "application/json")
-        self.send_header("Content-Length", str(len(b))); self.send_header("Cache-Control", "no-store"); self.end_headers(); self.wfile.write(b)
-
-    def _proxy(self, name, body=None):
-        if not BACKEND:
-            return self._json(503, {"error": "backend not configured"})
-        h = {"Content-Type": "application/json", "X-Backend-Token": BK_TOKEN}
-        if HF_AUTH:
-            h["Authorization"] = "Bearer " + HF_AUTH
-        req = urllib.request.Request(BACKEND + "/api/" + name, data=body, headers=h, method="POST" if body is not None else "GET")
-        try:
-            with urllib.request.urlopen(req, timeout=8 if name == "health" else TIMEOUT) as r:
-                data, code = r.read(), r.status
-            if not data.lstrip().startswith(b"{"):
-                return self._json(503, {"error": "backend starting"})
-            self.send_response(code); self.send_header("Content-Type", "application/json")
-            self.send_header("Content-Length", str(len(data))); self.send_header("Cache-Control", "no-store"); self.end_headers(); self.wfile.write(data)
-        except urllib.error.HTTPError as e:
-            self._json(503 if e.code >= 500 or e.code in (401, 403, 404) else e.code, {"error": "backend unavailable"})
-        except Exception:  # noqa: BLE001
-            self._json(503, {"error": "backend unavailable"})
-
-    def do_GET(self):
-        p = self.path.split("?")[0]
-        if p.startswith("/api/ag/"):
-            name = p[8:]
-            return self._proxy(name) if name in GET_OK else self._json(404, {"error": "not found"})
-        f = "index.html" if p == "/" else p.lstrip("/")
-        if f not in STATIC:
-            return self._json(404, {"error": "not found"})
-        self.path = "/" + f
-        return super().do_GET()
-
-    def do_POST(self):
-        p = self.path.split("?")[0]
-        name = p[8:] if p.startswith("/api/ag/") else ""
-        if name not in POST_OK:
-            return self._json(404, {"error": "not found"})
-        ip = (self.headers.get("X-Forwarded-For") or self.client_address[0]).split(",")[0].strip()
-        if limited(ip):
-            return self._json(429, {"error": "rate limit"})
-        n = int(self.headers.get("Content-Length", 0) or 0)
-        if n > 4096:
-            return self._json(413, {"error": "too large"})
-        try:
-            req = json.loads(self.rfile.read(n) or b"{}")
-        except ValueError:
-            return self._json(400, {"error": "bad json"})
-        if "message" in req:
-            req["message"] = str(req["message"])[:500]
-        self._proxy(name, json.dumps(req).encode())
-
-    def end_headers(self):
-        self.send_header("X-Content-Type-Options", "nosniff")
-        super().end_headers()
-
-    def log_message(self, *a):
-        pass
-
-
-if __name__ == "__main__":
-    os.chdir(os.path.dirname(os.path.abspath(__file__)))
-    ThreadingHTTPServer(("0.0.0.0", int(os.environ.get("PORT", "7860"))), H).serve_forever()
+"""ViKernel Desktop Space server: serves the static desktop and proxies /api/ag/* server-side
+to a private Agent Guard backend. Backend credentials live in Space secrets and never reach the browser."""
+import json, os, time, urllib.error, urllib.request
+from http.server import SimpleHTTPRequestHandler, ThreadingHTTPServer
+
+BACKEND = os.environ.get("BACKEND_URL", "").rstrip("/")
+BACKEND2 = os.environ.get("BACKEND2_URL", "").rstrip("/")  # in-weights graft backend (GPU)
+HF_AUTH = os.environ.get("BACKEND_HF_TOKEN", "")
+BK_TOKEN = os.environ.get("BACKEND_TOKEN", "")
+RATE = int(os.environ.get("RATE_PER_10MIN", "20"))
+TIMEOUT = float(os.environ.get("BACKEND_TIMEOUT", "300"))
+GET_OK = {"health", "status", "audit"}
+POST_OK = {"chat", "scenario", "kill", "resume", "tamper", "reset"}
+STATIC = {"index.html", "app.js", "ag.js", "i18n.js", "demo_data.js", "styles.css", "ag_replay.json"}
+HITS = {}
+
+
+def limited(ip):
+    now = time.time()
+    h = [x for x in HITS.get(ip, []) if now - x < 600]
+    full = len(h) >= RATE
+    if not full:
+        h.append(now)
+    HITS[ip] = h
+    if len(HITS) > 5000:
+        HITS.clear()
+    return full
+
+
+class H(SimpleHTTPRequestHandler):
+    def _json(self, code, obj):
+        b = json.dumps(obj).encode()
+        self.send_response(code); self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(b))); self.send_header("Cache-Control", "no-store"); self.end_headers(); self.wfile.write(b)
+
+    def _proxy(self, name, body=None, base=None):
+        base = base or BACKEND
+        if not base:
+            return self._json(503, {"error": "backend not configured"})
+        h = {"Content-Type": "application/json", "X-Backend-Token": BK_TOKEN}
+        if HF_AUTH:
+            h["Authorization"] = "Bearer " + HF_AUTH
+        req = urllib.request.Request(base + "/api/" + name, data=body, headers=h, method="POST" if body is not None else "GET")
+        try:
+            with urllib.request.urlopen(req, timeout=8 if name == "health" else TIMEOUT) as r:
+                data, code = r.read(), r.status
+            if not data.lstrip().startswith(b"{"):
+                return self._json(503, {"error": "backend starting"})
+            self.send_response(code); self.send_header("Content-Type", "application/json")
+            self.send_header("Content-Length", str(len(data))); self.send_header("Cache-Control", "no-store"); self.end_headers(); self.wfile.write(data)
+        except urllib.error.HTTPError as e:
+            self._json(503 if e.code >= 500 or e.code in (401, 403, 404) else e.code, {"error": "backend unavailable"})
+        except Exception:  # noqa: BLE001
+            self._json(503, {"error": "backend unavailable"})
+
+    def do_GET(self):
+        p = self.path.split("?")[0]
+        if p.startswith("/api/ag/") or p.startswith("/api/agx/"):
+            x = p.startswith("/api/agx/")
+            name = p[9:] if x else p[8:]
+            return self._proxy(name, base=BACKEND2 if x else None) if name in GET_OK else self._json(404, {"error": "not found"})
+        f = "index.html" if p == "/" else p.lstrip("/")
+        if f not in STATIC:
+            return self._json(404, {"error": "not found"})
+        self.path = "/" + f
+        return super().do_GET()
+
+    def do_POST(self):
+        p = self.path.split("?")[0]
+        x = p.startswith("/api/agx/")
+        name = p[9:] if x else (p[8:] if p.startswith("/api/ag/") else "")
+        if name not in POST_OK:
+            return self._json(404, {"error": "not found"})
+        ip = (self.headers.get("X-Forwarded-For") or self.client_address[0]).split(",")[0].strip()
+        if limited(ip):
+            return self._json(429, {"error": "rate limit"})
+        n = int(self.headers.get("Content-Length", 0) or 0)
+        if n > 4096:
+            return self._json(413, {"error": "too large"})
+        try:
+            req = json.loads(self.rfile.read(n) or b"{}")
+        except ValueError:
+            return self._json(400, {"error": "bad json"})
+        if "message" in req:
+            req["message"] = str(req["message"])[:500]
+        if x and not BACKEND2:
+            return self._json(503, {"error": "in-weights backend not configured"})
+        self._proxy(name, json.dumps(req).encode(), base=BACKEND2 if x else None)
+
+    def end_headers(self):
+        self.send_header("X-Content-Type-Options", "nosniff")
+        super().end_headers()
+
+    def log_message(self, *a):
+        pass
+
+
+if __name__ == "__main__":
+    os.chdir(os.path.dirname(os.path.abspath(__file__)))
+    ThreadingHTTPServer(("0.0.0.0", int(os.environ.get("PORT", "7860"))), H).serve_forever()