diff --git a/the-matrix/js/edge-worker-client.js b/the-matrix/js/edge-worker-client.js
index 5338229..eaa680f 100644
--- a/the-matrix/js/edge-worker-client.js
+++ b/the-matrix/js/edge-worker-client.js
@@ -2,13 +2,24 @@
  * edge-worker-client.js — Main-thread proxy for the edge-worker Web Worker.
  *
  * Spawns js/edge-worker.js as a module Worker and exposes:
- *   classify(text)   → Promise<{ label:'local'|'server', score, reason, localReply? }>
+ *   classify(text)   → Promise<{
+ *                         complexity: 'trivial'|'moderate'|'complex',
+ *                         score:      number,
+ *                         reason:     string,
+ *                         localReply?: string   // only when complexity === 'trivial'
+ *                       }>
  *   sentiment(text)  → Promise<{ label:'POSITIVE'|'NEGATIVE'|'NEUTRAL', score }>
  *   onReady(fn)      → register a callback fired when models finish loading
  *   isReady()        → boolean — true once both models are warm
+ *   warmup()         → start the worker early so first classify() is fast
  *
- * If Web Workers are unavailable (SSR / old browser), all calls fall back to
- * the naive "route to server" path so the app remains functional.
+ * Complexity tiers (set by the worker):
+ *   trivial  — greeting/small-talk; answered locally, 0 sats, no server call
+ *   moderate — simple question; show cost preview, route to server
+ *   complex  — technical/creative/code; always priced, show cost preview
+ *
+ * If Web Workers are unavailable (SSR / old browser), all calls fall back
+ * gracefully: classify → { complexity:'moderate', ... } so the app still works.
  */
 
 let _worker = null;
@@ -61,7 +72,8 @@ function _init() {
 
 function _fallback(type) {
   if (type === 'sentiment') return { label: 'NEUTRAL', score: 0.5 };
-  return { label: 'server', score: 0, reason: 'worker-unavailable' };
+  // classify fallback: moderate keeps the UI functional (shows estimate, routes to server)
+  return { complexity: 'moderate', score: 0, reason: 'worker-unavailable' };
 }
 
 function _send(type, text) {
diff --git a/the-matrix/js/edge-worker.js b/the-matrix/js/edge-worker.js
index fd5e178..e05ce2c 100644
--- a/the-matrix/js/edge-worker.js
+++ b/the-matrix/js/edge-worker.js
@@ -29,10 +29,17 @@ import { pipeline, env } from '@xenova/transformers';
 env.useBrowserCache = true;     // cache model weights via browser Cache API
 env.allowLocalModels = false;   // no filesystem — browser-only environment
 
-const LOCAL_LABELS  = ['greeting', 'small-talk', 'simple-question'];
-const SERVER_LABELS = ['technical-task', 'creative-work', 'complex-question', 'code-request'];
-const ALL_LABELS    = [...LOCAL_LABELS, ...SERVER_LABELS];
-const LOCAL_THRESHOLD = 0.55;
+// Classification labels → complexity tier mapping
+// trivial  — handled locally, no server call, no sats
+// moderate — real request but may be free-tier; show cost preview before send
+// complex  — substantive work; always priced; show cost preview before send
+const TRIVIAL_LABELS  = ['greeting', 'small-talk'];
+const MODERATE_LABELS = ['simple-question'];
+const COMPLEX_LABELS  = ['technical-task', 'creative-work', 'complex-question', 'code-request'];
+const ALL_LABELS      = [...TRIVIAL_LABELS, ...MODERATE_LABELS, ...COMPLEX_LABELS];
+
+const TRIVIAL_THRESHOLD  = 0.55; // minimum score to call trivial "trivial"
+const MODERATE_THRESHOLD = 0.40; // below this → upgrade to complex (model is uncertain)
 
 const LOCAL_REPLIES = [
   "Greetings, traveller! Ask me something arcane and I shall conjure wisdom from the ether.",
@@ -64,31 +71,55 @@ async function _loadModels() {
 
 // ── Handlers ──────────────────────────────────────────────────────────────────
 
+/**
+ * _classify — returns the complexity tier of the input text.
+ *
+ * Returns:
+ *   {
+ *     complexity: 'trivial' | 'moderate' | 'complex',
+ *     score:      number,        // top-label confidence
+ *     reason:     string,        // winning label name or heuristic name
+ *     localReply?: string,       // only present when complexity === 'trivial'
+ *   }
+ *
+ * Complexity tiers:
+ *   trivial  — greeting/small-talk with high confidence; answered locally, 0 sats
+ *   moderate — simple question or uncertain classification; show cost preview, route to server
+ *   complex  — technical / creative / code work; always priced, show cost preview
+ */
 async function _classify(text) {
   const trimmed = text.trim();
 
+  // Fast path: single-word/short greetings are trivially local
   if (_isGreeting(trimmed)) {
-    return { label: 'local', score: 0.99, reason: 'greeting-heuristic', localReply: _randomReply() };
+    return { complexity: 'trivial', score: 0.99, reason: 'greeting-heuristic', localReply: _randomReply() };
   }
 
+  // If model hasn't loaded yet, default to moderate (show estimate, let server decide)
   if (!_classifier) {
-    return { label: 'server', score: 0, reason: 'model-unavailable' };
+    return { complexity: 'moderate', score: 0, reason: 'model-unavailable' };
   }
 
   try {
-    const result = await _classifier(trimmed, ALL_LABELS, { multi_label: false });
+    const result   = await _classifier(trimmed, ALL_LABELS, { multi_label: false });
     const topLabel = result.labels[0];
     const topScore = result.scores[0];
-    const isLocal  = LOCAL_LABELS.includes(topLabel) && topScore >= LOCAL_THRESHOLD;
 
-    return {
-      label:  isLocal ? 'local' : 'server',
-      score:  topScore,
-      reason: topLabel,
-      ...(isLocal ? { localReply: _randomReply() } : {}),
-    };
+    if (TRIVIAL_LABELS.includes(topLabel) && topScore >= TRIVIAL_THRESHOLD) {
+      // High-confidence trivial: small-talk or greeting  → answer locally
+      return { complexity: 'trivial', score: topScore, reason: topLabel, localReply: _randomReply() };
+    }
+
+    if (COMPLEX_LABELS.includes(topLabel) || topScore < MODERATE_THRESHOLD) {
+      // Explicitly complex label, or model is uncertain (score too low to trust) → complex
+      return { complexity: 'complex', score: topScore, reason: topLabel };
+    }
+
+    // Middle ground: simple-question or trivial label with moderate confidence → moderate
+    return { complexity: 'moderate', score: topScore, reason: topLabel };
+
   } catch (err) {
-    return { label: 'server', score: 0, reason: 'classify-error', error: String(err) };
+    return { complexity: 'moderate', score: 0, reason: 'classify-error', error: String(err) };
   }
 }
 
diff --git a/the-matrix/js/nostr-identity.js b/the-matrix/js/nostr-identity.js
index a71291e..b17b78a 100644
--- a/the-matrix/js/nostr-identity.js
+++ b/the-matrix/js/nostr-identity.js
@@ -30,6 +30,7 @@ let _pubkey    = null;  // hex
 let _token     = null;  // nostr_token string from server
 let _tokenExp  = 0;     // unix ms when token was fetched
 let _useNip07  = false; // true if window.nostr is available
+let _canSign   = false; // true if we have a signing capability (NIP-07 or local privkey)
 let _identityPromptShown = false;
 
 // ── Init ──────────────────────────────────────────────────────────────────────
@@ -40,6 +41,7 @@ export async function initNostrIdentity(apiBase = '/api') {
     try {
       _pubkey   = await window.nostr.getPublicKey();
       _useNip07 = true;
+      _canSign  = true;
       console.info('[nostr] Using NIP-07 extension, pubkey:', _pubkey.slice(0, 8) + '…');
     } catch (err) {
       console.warn('[nostr] NIP-07 getPublicKey failed, will use local keypair', err);
@@ -47,24 +49,33 @@ export async function initNostrIdentity(apiBase = '/api') {
     }
   }
 
-  // Try restoring an existing local keypair (consented previously)
+  // Try restoring / discovering an existing keypair
   if (!_pubkey) {
     _pubkey = _loadKeypair();
     if (_pubkey) {
-      console.info('[nostr] Restored local keypair, pubkey:', _pubkey.slice(0, 8) + '…');
+      // Check if we actually have a privkey for signing (npub-only discovery gives pubkey but no privkey)
+      _canSign = !!_getPrivkeyBytes();
+      if (_canSign) {
+        console.info('[nostr] Restored local keypair (with signing), pubkey:', _pubkey.slice(0, 8) + '…');
+      } else {
+        console.info('[nostr] Discovered pubkey (view-only, no privkey), pubkey:', _pubkey.slice(0, 8) + '…');
+      }
     }
   }
 
   // Try to restore cached token
   _loadCachedToken();
 
-  // If we have a pubkey but no valid token, authenticate now
-  if (_pubkey && !_isTokenValid()) {
+  // If we have signing capability and no valid token, authenticate now
+  if (_pubkey && _canSign && !_isTokenValid()) {
     await refreshToken(apiBase);
   }
 
-  // If no identity at all, show the opt-in prompt so the user can choose
-  if (!_pubkey) {
+  // Show the opt-in prompt if:
+  //   a) No identity at all — user can generate a keypair or connect NIP-07
+  //   b) Have a pubkey but no signing capability (npub-only discovery) —
+  //      user should be offered a way to add signing (generate new key or NIP-07)
+  if (!_pubkey || !_canSign) {
     _scheduleIdentityPrompt(apiBase);
   }
 }
@@ -185,12 +196,14 @@ export function showIdentityPrompt(apiBase = '/api') {
       try {
         _pubkey   = await window.nostr.getPublicKey();
         _useNip07 = true;
+        _canSign  = true;
       } catch { return; }
     } else {
       // Generate + store keypair (user consented by clicking)
-      _pubkey = _generateAndSaveKeypair();
+      _pubkey  = _generateAndSaveKeypair();
+      _canSign = true;
     }
-    if (_pubkey) {
+    if (_pubkey && _canSign) {
       await refreshToken(apiBase);
       _updateIdentityHUD();
     }
diff --git a/the-matrix/js/session.js b/the-matrix/js/session.js
index 29c2cd3..bfaedf7 100644
--- a/the-matrix/js/session.js
+++ b/the-matrix/js/session.js
@@ -278,7 +278,9 @@ function _startDepositPolling() {
   async function poll() {
     if (!_sessionId) return;
     try {
-      const res  = await fetch(`${API}/sessions/${_sessionId}`);
+      const nostrToken = await getOrRefreshToken('/api');
+      const pollHeaders = nostrToken ? { 'X-Nostr-Token': nostrToken } : {};
+      const res  = await fetch(`${API}/sessions/${_sessionId}`, { headers: pollHeaders });
       const data = await res.json();
 
       if (data.state === 'active') {
@@ -322,12 +324,16 @@ async function _createTopup() {
   _btn('session-topup-create-btn', true);
 
   try {
+    const nostrToken = await getOrRefreshToken('/api');
+    const topupHeaders = {
+      'Content-Type': 'application/json',
+      'Authorization': `Bearer ${_macaroon}`,
+    };
+    if (nostrToken) topupHeaders['X-Nostr-Token'] = nostrToken;
+
     const res  = await fetch(`${API}/sessions/${_sessionId}/topup`, {
       method:  'POST',
-      headers: {
-        'Content-Type': 'application/json',
-        'Authorization': `Bearer ${_macaroon}`,
-      },
+      headers: topupHeaders,
       body: JSON.stringify({ amount_sats: _topupSats }),
     });
     const data = await res.json();
@@ -388,7 +394,9 @@ function _startTopupPolling() {
   async function poll() {
     if (!_sessionId) return;
     try {
-      const res  = await fetch(`${API}/sessions/${_sessionId}`);
+      const nostrToken = await getOrRefreshToken('/api');
+      const pollHeaders = nostrToken ? { 'X-Nostr-Token': nostrToken } : {};
+      const res  = await fetch(`${API}/sessions/${_sessionId}`, { headers: pollHeaders });
       const data = await res.json();
 
       if (data.balanceSats > prevBalance || data.state === 'active') {
@@ -421,7 +429,9 @@ async function _tryRestore() {
     if (!sessionId || !macaroon) return;
 
     // Validate the session is still live
-    const res  = await fetch(`${API}/sessions/${sessionId}`);
+    const nostrToken = await getOrRefreshToken('/api');
+    const restoreHeaders = nostrToken ? { 'X-Nostr-Token': nostrToken } : {};
+    const res  = await fetch(`${API}/sessions/${sessionId}`, { headers: restoreHeaders });
     if (!res.ok) { localStorage.removeItem(LS_KEY); return; }
     const data = await res.json();
 
diff --git a/the-matrix/js/ui.js b/the-matrix/js/ui.js
index c824102..dc5c700 100644
--- a/the-matrix/js/ui.js
+++ b/the-matrix/js/ui.js
@@ -150,10 +150,11 @@ function initInputBar() {
     }
 
     // ── Edge triage — classify text in the Web Worker ─────────────────────────
+    // Worker returns { complexity: 'trivial'|'moderate'|'complex', score, reason, localReply? }
     const cls = await classify(text);
 
-    if (cls.label === 'local' && cls.localReply) {
-      // Trivial/conversational — answer locally, no server round-trip
+    if (cls.complexity === 'trivial' && cls.localReply) {
+      // Greeting / small-talk → answer locally, 0 sats, no network call
       appendSystemMessage(`you: ${text}`);
       appendSystemMessage(`Timmy [local]: ${cls.localReply}`);
       _showCostPreview('answered locally ⚡ 0 sats', '#44dd88');
@@ -161,7 +162,15 @@ function initInputBar() {
       return;
     }
 
-    // Substantive — route to server via WebSocket
+    // moderate or complex — fetch cost estimate (driven by complexity outcome),
+    // then route to server. The preview is already shown via debounce on input,
+    // but we refresh it now with the actual classification context.
+    if (cls.complexity === 'moderate' || cls.complexity === 'complex') {
+      // Fire estimate fetch; don't await — let it update the badge async while WS is in flight
+      _fetchEstimate(text);
+    }
+
+    // Route to server via WebSocket
     sendVisitorMessage(text);
     appendSystemMessage(`you: ${text}`);
   }