Ship Assistent 0.10.8 cheap-bug sweep for RU intent, VRAM park, and identity.

Silent Generate, pack selection, and loading UI no longer trip on Cyrillic or fake GPU loads; persona title is the assistant, not the user.

Co-authored-by: Cursor <cursoragent@cursor.com>

Assets/assistent.js

@@ -14,6 +14,8 @@
     const LS_AUTO_GENERATE = 'swarm_assistent_auto_generate';
     const LS_AUTO_CRITIQUE = 'swarm_assistent_auto_critique';
     const LS_AUTO_DOWNLOAD = 'swarm_assistent_auto_download';
+    /** When '1', unload chat LLM before Generate (frees VRAM; VL reload can take 1–2 min). Default off. */
+    const LS_PARK_LLM = 'swarm_assistent_park_llm';
     const LS_PANE_WIDTH = 'swarm_assistent_pane_width';
     const LS_WELCOMED = 'swarm_assistent_welcomed';
     const LS_TASTE = 'swarm_assistent_taste';
@@ -163,6 +165,7 @@
         chatsSearchHits: null,
         slashIndex: 0,
         llmParked: false,
+        expectColdLoad: false,
         memoryRows: [],
         userPrefs: [],
         settingsTab: 'behavior',
@@ -224,14 +227,17 @@
             return;
         }
         const elapsed = Date.now() - (state.busyStarted || Date.now());
+        // Only claim "Loading into GPU" when the model was parked / known cold.
+        // Otherwise a slow first token on an already-resident VL looks like a 2‑min reload.
         if (!state.gotDelta && (state.busyPhase === 'thinking' || state.busyPhase === 'waiting') && elapsed > 1600) {
-            state.busyPhase = 'loading';
+            state.busyPhase = state.llmParked || state.expectColdLoad ? 'loading' : 'waiting';
         }
         const model = modelShort($('sa_model')?.value);
         const labels = {
             encoding: 'Encoding image…',
-            waiting: 'Waiting for Ollama…',
+            waiting: 'Жду Ollama / первый токен…',
-            loading: `Loading ${model} into GPU… first load can take a minute`,
+            loading: `Загружаю ${model} в GPU… обычно 30–120 с после park`,
+            warming: `Возвращаю ${model} в GPU…`,
             thinking: 'Thinking…',
             streaming: 'Writing…',
             generating: 'Generating image…',
@@ -482,19 +488,42 @@
         return false;
     }
 
+    /** JS \\b/\\w are ASCII-only — use this for RU tokens. `alts` = regex alternatives without outer parens. */
+    function cyrTokenRe(alts) {
+        const boundary = '(^|[^0-9A-Za-z_А-Яа-яЁё])';
+        const end = '(?=$|[^0-9A-Za-z_А-Яа-яЁё])';
+        return new RegExp(`${boundary}(?:${alts})${end}`, 'i');
+    }
+
     function userTextMentionsParams(text) {
-        return /\b(steps?|cfg|seed|sigma|размер|aspect|ширин|высот|resolution|batch|турбо|turbo|raw)\b/i.test(String(text || ''));
+        const t = String(text || '');
+        if (/\b(steps?|cfg|seed|sigma|aspect|resolution|batch|turbo|raw)\b/i.test(t)) {
+            return true;
+        }
+        return cyrTokenRe('размер|ширин[а-яё]*|высот[а-яё]*|турбо').test(t);
     }
 
     function userAsksGenerate(text) {
-        const t = String(text || '');
+        const t = String(text || '').trim();
-        if (!t.trim()) {
+        if (!t) {
             return false;
         }
-        if (/^(gen|generate|go|рисуй|давай|ещё|еще)\s*[!.…]*$/i.test(t.trim())) {
+        // Short imperatives only — do NOT treat bare «давай» as Generate (false positive on chat).
+        if (/^(gen|generate|go|рисуй|нарисуй)([!.…\s]|$)/i.test(t)) {
             return true;
         }
-        return /\b(сгенерируй|сгенерировать|генерируй|generate|нарисуй|перегенерируй|перерисуй|сделай\s+(картинк|изображен|фото)|run\s+generat|\/gen)\b/i.test(t);
+        if (/^(ещё|еще)(\s+раз)?([!.…\s]|$)/i.test(t)) {
+            return true;
+        }
+        // Do NOT use \b or \w — ASCII-only in JS; breaks «сделай картинку».
+        const letter = '[0-9A-Za-z_А-Яа-яЁё]';
+        const stem = `${letter}*`;
+        return cyrTokenRe(
+            'сгенерируй|сгенерировать|генерируй|generate|нарисуй|перегенерируй|перерисуй|'
+            + `сделай\\s+(картинк${stem}|изображен${stem}|фото${stem})|`
+            + `хочу\\s+(картинк${stem}|изображен${stem}|фото${stem})|`
+            + 'run\\s+generat|/gen',
+        ).test(t);
     }
 
     function rememberLastPatch(patch) {
@@ -1374,16 +1403,29 @@
         return `c_${Date.now().toString(36)}_${Math.random().toString(36).slice(2, 8)}`;
     }
 
+    function stripJsonFencesForHistory(content) {
+        return String(content || '')
+            .replace(/```(?:json)?\s*[\s\S]*?```/gi, '')
+            .replace(/\n{3,}/g, '\n\n')
+            .trim();
+    }
+
     function slimHistoryMessages(list) {
         return (list || [])
             .filter((m) => m && (m.role === 'user' || m.role === 'assistant') && !m.systemish)
             .slice(-MAX_CHAT_MSGS)
-            .map((m) => ({
+            .map((m) => {
-                role: m.role,
+                let content = String(m.content || '');
-                content: String(m.content || '').slice(0, 4000),
+                if (m.role === 'assistant') {
-                persona: m.persona || undefined,
+                    content = stripJsonFencesForHistory(content);
-                pack: m.pack || undefined,
+                }
-            }));
+                return {
+                    role: m.role,
+                    content: content.slice(0, 4000),
+                    persona: m.persona || undefined,
+                    pack: m.pack || undefined,
+                };
+            });
     }
 
     function titleFromMessages(messages) {
@@ -2799,17 +2841,24 @@
         if (!obj || typeof obj !== 'object') {
             return false;
         }
+        if (isCardObject(obj)) {
+            return false;
+        }
         return FALLBACK_PATCH_KEYS.some((k) => obj[k] !== undefined && obj[k] !== null);
     }
 
     function isCardObject(obj) {
+        if (window.SA && typeof SA.isCardObject === 'function') {
+            return SA.isCardObject(obj);
+        }
         if (!obj || typeof obj !== 'object') {
             return false;
         }
         // Prefer card shape over gen patch when both could match.
         const cardish = !!(obj.kind || obj.triggers || obj.when || obj.prompt_hint);
         const genish = !!(obj.prompt != null || obj.loras || obj.actions || obj.width || obj.height
-            || obj.steps || obj.cfg || obj.aspect || obj.seed != null);
+            || obj.steps || obj.cfg || obj.aspect || obj.seed != null
+            || obj.search_query || obj.civitai_query || obj.look_at || obj.controls);
         if (cardish && !genish && (obj.name || obj.triggers || obj.when)) {
             return true;
         }
@@ -2992,23 +3041,31 @@
         if (!t.trim()) {
             return null;
         }
-        if (/\b(поправь|исправь|перепиши|улучши|fix\s+it|make\s+it\s+better|rewrite)\b/i.test(t)) {
+        if (cyrTokenRe('поправь|исправь|перепиши|улучши').test(t)
+            || /\b(fix\s+it|make\s+it\s+better|rewrite)\b/i.test(t)) {
             return 'write_prompt';
         }
-        if (/\b(опиши\s+реф|опиши\s+изображ|prompt\s+from\s+image|describe\s+(this|the|ref|image)|reverse\s*prompt)\b/i.test(t)
+        if (cyrTokenRe('опиши\\s+(реф|изображ[а-яё]*|этот|эту|картинк[а-яё]*|референс)').test(t)
+            || /\b(prompt\s+from\s+image|describe\s+(this|the|ref|image)|reverse\s*prompt)\b/i.test(t)
             || /опиши\s+(этот|эту|картинк|референс)/i.test(t)) {
             return 'describe_ref';
         }
-        if (/\b(critique|что\s+не\s+так|посмотри|смотри|разбери|критик)\b/i.test(t)) {
+        // Bare «посмотри/смотри» is casual chat — only critique when aimed at a result/frame.
+        if (/\b(critique|criticize)\b/i.test(t)
+            || cyrTokenRe('критик[а-яё]*|что\\s+не\\s+так|разбери').test(t)
+            || /(?:^|[^а-яёa-z0-9_])(посмотри|смотри)\s+(на\s+)?(результат|картинк[а-яё]*|изображен[а-яё]*|кадр|ген)/i.test(t)) {
             return 'critique_image';
         }
-        if (/\b(inpaint|замажь|закрась|руки|лицо|маск|mask|img2img|init\s*image)\b/i.test(t)) {
+        if (/\b(inpaint|mask|img2img)\b/i.test(t)
+            || cyrTokenRe('замажь|закрась|руки|лицо|маск[а-яё]*').test(t)
+            || /init\s*image/i.test(t)) {
             return 'inpaint_edit';
         }
         if (userTextMentionsParams(t)) {
-            return 'fix_params';
+            return 'form_params';
         }
-        if (/\b(сцен|moodboard|атмосфер|compose|scene|мизансцен)\b/i.test(t)) {
+        if (/\b(moodboard|compose|scene)\b/i.test(t)
+            || cyrTokenRe('сцен[а-яё]*|атмосфер[а-яё]*|мизансцен[а-яё]*').test(t)) {
             return 'compose_scene';
         }
         return 'write_prompt';
@@ -3402,11 +3459,15 @@
         return false;
     }
 
+    function shouldParkLlmBeforeGen() {
+        return !!$('sa_park_llm')?.checked;
+    }
+
     /** Unloads the chat model from VRAM so Krea 2 gets the whole GPU. Never touches the embed model. */
     function parkLlm() {
         return new Promise((resolve) => {
             const model = $('sa_model')?.value;
-            if (!model || state.llmParked || typeof genericRequest !== 'function') {
+            if (!shouldParkLlmBeforeGen() || !model || state.llmParked || typeof genericRequest !== 'function') {
                 resolve(false);
                 return;
             }
@@ -3419,6 +3480,7 @@
                 settled = true;
                 if (ok) {
                     state.llmParked = true;
+                    state.expectColdLoad = true;
                 }
                 resolve(!!ok);
             };
@@ -3427,15 +3489,31 @@
         });
     }
 
-    /** Fire-and-forget re-load of the chat model once the user is back in the chat. */
+    /** Re-load chat model after park. Returns a Promise (await after Generate so the next Send is warm). */
     function warmLlm() {
-        const model = $('sa_model')?.value;
+        return new Promise((resolve) => {
-        if (!model || !state.llmParked || typeof genericRequest !== 'function') {
+            const model = $('sa_model')?.value;
-            return;
+            if (!model || !state.llmParked || typeof genericRequest !== 'function') {
-        }
+                resolve(false);
-        state.llmParked = false;
+                return;
-        const baseUrl = $('sa_base_url')?.value || 'http://127.0.0.1:11434';
+            }
-        genericRequest('AssistentWarmLlm', { baseUrl, model }, () => {}, 0, () => {});
+            const baseUrl = $('sa_base_url')?.value || 'http://127.0.0.1:11434';
+            let settled = false;
+            const finish = (ok) => {
+                if (settled) {
+                    return;
+                }
+                settled = true;
+                state.llmParked = false;
+                if (ok) {
+                    state.expectColdLoad = false;
+                }
+                resolve(!!ok);
+            };
+            // VL 7B cold-load can exceed a minute — don't time out the flag early.
+            setTimeout(() => finish(false), 180000);
+            genericRequest('AssistentWarmLlm', { baseUrl, model }, () => finish(true), 0, () => finish(false));
+        });
     }
 
     function cancelWaitForNewImage() {
@@ -3528,9 +3606,11 @@
             return null;
         }
         const prev = findCurrentGenerateSrc();
-        startBusyUi('parking');
+        if (shouldParkLlmBeforeGen()) {
-        setStatus('Освобождаю VRAM…');
+            startBusyUi('parking');
-        await parkLlm();
+            setStatus('Освобождаю VRAM…');
+            await parkLlm();
+        }
         setStatus('Генерация…');
         startBusyUi('generating');
         state.generating = true;
@@ -3545,14 +3625,16 @@
         const src = await waitForNewImage(prev);
         state.generating = false;
         setInterruptVisible(state.busy);
+        // Auto-critique / next Send load the model themselves — skip warm if critique will run.
+        const paneVisible = !!document.getElementById('swarm_assistent_root')?.offsetParent;
+        if (state.llmParked && state.view === 'chat' && paneVisible && !$('sa_auto_critique')?.checked) {
+            startBusyUi('warming');
+            setStatus('Возвращаю LLM в GPU…');
+            await warmLlm();
+        }
         if (!state.busy) {
             stopBusyUi(src ? 'Generate готов' : 'Generate завершён (новое изображение не найдено)');
         }
-        // Auto-critique loads the model itself on the next request — don't pay for it twice.
-        const paneVisible = !!document.getElementById('swarm_assistent_root')?.offsetParent;
-        if (state.view === 'chat' && paneVisible && !$('sa_auto_critique')?.checked) {
-            warmLlm();
-        }
         if (src) {
             const gen = generateSlot();
             if (gen) {
@@ -3777,6 +3859,7 @@
                 setAssistantBody(state.streamEl, '');
             }
             state.gotDelta = true;
+            state.expectColdLoad = false;
             if (state.busyPhase !== 'refining') {
                 setBusyPhase('streaming');
             }
@@ -4161,6 +4244,7 @@
         const autoGen = localStorage.getItem(LS_AUTO_GENERATE);
         const autoCrit = localStorage.getItem(LS_AUTO_CRITIQUE);
         const autoDl = localStorage.getItem(LS_AUTO_DOWNLOAD);
+        const parkLlm = localStorage.getItem(LS_PARK_LLM);
         const paneW = localStorage.getItem(LS_PANE_WIDTH);
         if (base && $('sa_base_url')) {
             $('sa_base_url').value = base;
@@ -4186,6 +4270,10 @@
         if ($('sa_auto_download') && autoDl != null) {
             $('sa_auto_download').checked = autoDl === '1';
         }
+        // Default OFF — parking a VL 7B before every Generate caused 1–2 min reloads.
+        if ($('sa_park_llm')) {
+            $('sa_park_llm').checked = parkLlm === '1';
+        }
         if (model) {
             state.preferredModel = model;
         }
@@ -4214,6 +4302,7 @@
             auto_generate: !!$('sa_auto_generate')?.checked,
             auto_critique: !!$('sa_auto_critique')?.checked,
             auto_download: !!$('sa_auto_download')?.checked,
+            park_llm: !!$('sa_park_llm')?.checked,
             pane_width: localStorage.getItem(LS_PANE_WIDTH) || '',
             embed_model: $('sa_embed_model')?.value || state.preferredEmbed || '',
             base_url: $('sa_base_url')?.value || '',
@@ -4267,6 +4356,7 @@
             ['auto_generate', LS_AUTO_GENERATE, 'sa_auto_generate'],
             ['auto_critique', LS_AUTO_CRITIQUE, 'sa_auto_critique'],
             ['auto_download', LS_AUTO_DOWNLOAD, 'sa_auto_download'],
+            ['park_llm', LS_PARK_LLM, 'sa_park_llm'],
         ]) {
             if (ui[key] == null || localStorage.getItem(lsKey) != null) {
                 continue;
@@ -4299,6 +4389,7 @@
         localStorage.setItem(LS_AUTO_GENERATE, $('sa_auto_generate')?.checked ? '1' : '0');
         localStorage.setItem(LS_AUTO_CRITIQUE, $('sa_auto_critique')?.checked ? '1' : '0');
         localStorage.setItem(LS_AUTO_DOWNLOAD, $('sa_auto_download')?.checked ? '1' : '0');
+        localStorage.setItem(LS_PARK_LLM, $('sa_park_llm')?.checked ? '1' : '0');
         persistServerSettings();
         saveUiStateToDisk();
     }
@@ -4418,11 +4509,20 @@
     }
 
     let controlSaveTimer = null;
+    let controlsPointerDown = false;
+    let pendingControlsRender = null;
+
     function renderPersonaControls(schema, values) {
         const box = $('sa_persona_controls');
         if (!box) {
             return;
         }
+        // Don't rebuild DOM while the user is dragging — that snaps the thumb back.
+        if (controlsPointerDown) {
+            pendingControlsRender = { schema, values };
+            return;
+        }
+        pendingControlsRender = null;
         box.innerHTML = '';
         const keys = schema && typeof schema === 'object' ? Object.keys(schema) : [];
         if (!keys.length) {
@@ -4471,16 +4571,40 @@
             const valEl = document.createElement('span');
             valEl.className = 'sa-control-val';
             valEl.textContent = fmt(cur);
-            const onInput = () => {
+            const applyLocal = (v) => {
-                const v = Number(input.value);
                 valEl.textContent = fmt(v);
+                if (state.config) {
+                    state.config.control_values = { ...(state.config.control_values || {}), [id]: v };
+                }
+                if (state.exact) {
+                    state.exact.controls = { ...(state.exact.controls || {}), [id]: v };
+                }
+            };
+            input.addEventListener('pointerdown', () => {
+                controlsPointerDown = true;
+            });
+            const endPointer = () => {
+                controlsPointerDown = false;
+                if (pendingControlsRender) {
+                    const pending = pendingControlsRender;
+                    pendingControlsRender = null;
+                    renderPersonaControls(pending.schema, pending.values);
+                }
+            };
+            input.addEventListener('pointerup', endPointer);
+            input.addEventListener('pointercancel', endPointer);
+            // Live label while dragging; persist only on release (change) so mid-drag saves cannot snap back.
+            input.addEventListener('input', () => {
+                applyLocal(Number(input.value));
+            });
+            input.addEventListener('change', () => {
+                const v = Number(input.value);
+                applyLocal(v);
                 if (controlSaveTimer) {
                     clearTimeout(controlSaveTimer);
                 }
-                controlSaveTimer = setTimeout(() => savePersonaControls({ [id]: v }), 350);
+                controlSaveTimer = setTimeout(() => savePersonaControls({ [id]: v }), 50);
-            };
+            });
-            input.addEventListener('input', onInput);
-            input.addEventListener('change', onInput);
             row.appendChild(lab);
             row.appendChild(input);
             row.appendChild(valEl);
@@ -4509,11 +4633,18 @@
         const max = Number(schema.horny.max ?? 100);
         const cur = getControlValue('horny', Number(schema.horny.default ?? 35));
         const next = Math.max(min, Math.min(max, cur - 30));
-        savePersonaControls({ horny: next });
+        const values = {
-        renderPersonaControls(schema, {
             ...(state.config?.control_values || state.exact?.controls || {}),
             horny: next,
-        });
+        };
+        if (state.config) {
+            state.config.control_values = values;
+        }
+        if (state.exact) {
+            state.exact.controls = values;
+        }
+        renderPersonaControls(schema, values);
+        savePersonaControls({ horny: next });
         appendSystemNote(`Хорни: ${Math.round(cur)}% → ${Math.round(next)}% (−30)`);
         setStatus(`/остынь → ${Math.round(next)}%`);
     }
@@ -4544,6 +4675,15 @@
         if (typeof genericRequest !== 'function') {
             return;
         }
+        // Optimistic local merge so UI / next chat see the new values immediately.
+        if (partial && typeof partial === 'object') {
+            if (state.config) {
+                state.config.control_values = { ...(state.config.control_values || {}), ...partial };
+            }
+            if (state.exact) {
+                state.exact.controls = { ...(state.exact.controls || {}), ...partial };
+            }
+        }
         genericRequest(
             'AssistentSaveControls',
             { persona, controls: partial || {} },
@@ -4558,8 +4698,10 @@
                         state.exact.controls = data.control_values;
                     }
                 }
-                if (data?.controls) {
+                // Do not rebuild slider DOM here — that interrupts an in-progress drag and snaps values back.
-                    renderPersonaControls(data.controls, data.control_values || {});
+                // Update live inputs in place if present and not being dragged.
+                if (!controlsPointerDown && data?.control_values) {
+                    syncPersonaControlInputs(data.control_values);
                 }
             },
             0,
@@ -4567,6 +4709,30 @@
         );
     }
 
+    function syncPersonaControlInputs(values) {
+        const box = $('sa_persona_controls');
+        if (!box || !values || typeof values !== 'object') {
+            return;
+        }
+        box.querySelectorAll('input[data-control-id]').forEach((input) => {
+            const id = input.dataset.controlId;
+            if (values[id] == null) {
+                return;
+            }
+            const v = Number(values[id]);
+            if (!Number.isFinite(v) || input.value === String(v)) {
+                return;
+            }
+            input.value = String(v);
+            const valEl = input.parentElement?.querySelector('.sa-control-val');
+            if (valEl) {
+                const schema = state.config?.controls?.[id];
+                const asPercent = String(schema?.display || '').toLowerCase() === 'percent';
+                valEl.textContent = asPercent ? `${Math.round(v)}%` : Number(v).toFixed(2);
+            }
+        });
+    }
+
     async function deleteCurrentOverlayPersona() {
         const id = $('sa_persona')?.value;
         if (!id) {
@@ -6990,10 +7156,14 @@
 
         const chatEpoch = bumpChatEpoch();
         state.busy = true;
-        // Ollama reloads the model for this request (keep_alive 15m), so it is no longer parked.
+        // This request will keep_alive 15m; clear parked only after we know load started.
+        // Keep expectColdLoad so the UI can show a real GPU-load message if we just parked.
+        if (!state.llmParked) {
+            state.expectColdLoad = false;
+        }
         state.llmParked = false;
         setInterruptVisible(true);
-        startBusyUi('thinking');
+        startBusyUi(state.expectColdLoad ? 'loading' : 'thinking');
         saveSettings();
 
         // Always pull latest LoRA/checkpoint list before the LLM sees context
@@ -7079,7 +7249,13 @@
         // Refresh cards into context after prefetch
         const refreshed = collectLiveContext();
         context.model_cards = refreshed.model_cards;
-        const messages = state.history.slice(-historyMessageLimit()).map((m) => ({ role: m.role, content: m.content }));
+        const messages = state.history.slice(-historyMessageLimit()).map((m) => {
+            let content = String(m.content || '');
+            if (m.role === 'assistant') {
+                content = stripJsonFencesForHistory(content);
+            }
+            return { role: m.role, content: content.slice(0, 4000) };
+        });
         if (images && messages.length) {
             messages[messages.length - 1].images = images;
         }
@@ -7127,10 +7303,14 @@
                 if (chatEpoch !== state.chatEpoch) {
                     return;
                 }
-                state.busy = false;
+                // Keep busy while silent Apply+Generate is still running (generating flag).
-                setInterruptVisible(state.generating);
                 if (!state.generating) {
+                    state.busy = false;
+                    setInterruptVisible(false);
                     stopBusyUi('Готово');
+                } else {
+                    state.busy = false;
+                    setInterruptVisible(true);
                 }
             }
         };
@@ -7163,10 +7343,13 @@
                         return;
                     }
                     if (data.phase === 'waiting_ollama') {
-                        setBusyPhase('loading');
+                        // Server always emits this before /api/chat — not proof of a cold load.
+                        setBusyPhase(state.expectColdLoad ? 'loading' : 'waiting');
                         const label = state.streamEl?.querySelector('.sa-typing-label');
                         if (label) {
-                            label.textContent = `Загружаю ${modelShort(model)} в GPU…`;
+                            label.textContent = state.expectColdLoad
+                                ? `Загружаю ${modelShort(model)} в GPU…`
+                                : 'Думаю…';
                         }
                         return;
                     }
@@ -7748,6 +7931,7 @@
         $('sa_auto_generate')?.addEventListener('change', saveSettings);
         $('sa_auto_critique')?.addEventListener('change', saveSettings);
         $('sa_auto_download')?.addEventListener('change', saveSettings);
+        $('sa_park_llm')?.addEventListener('change', saveSettings);
 
         syncChipHighlight();
         setInterval(syncChipHighlight, 2500);

Assets/assistent.patch.js

@@ -24,11 +24,29 @@ window.SA = window.SA || {};
         return obj[key] !== undefined && obj[key] !== null;
     }
 
-    /** True when the object looks like a generation patch rather than arbitrary JSON. */
+    /** Model-card JSON (catalog_card) — must not be treated as a Generate patch. */
+    function isCardObject(obj) {
+        if (!obj || typeof obj !== 'object') {
+            return false;
+        }
+        const cardish = !!(obj.kind || obj.triggers || obj.when || obj.prompt_hint);
+        const genish = !!(obj.prompt != null || obj.loras || obj.actions || obj.width || obj.height
+            || obj.steps || obj.cfg || obj.aspect || obj.seed != null
+            || obj.search_query || obj.civitai_query || obj.look_at || obj.controls);
+        if (cardish && !genish && (obj.name || obj.triggers || obj.when)) {
+            return true;
+        }
+        return !!(obj.kind && obj.name && (obj.triggers || obj.when || obj.prompt_hint || obj.notes != null));
+    }
+
+    /** True when the object looks like a generation patch rather than a catalog card / arbitrary JSON. */
     function isPatchObject(obj) {
         if (!obj || typeof obj !== 'object') {
             return false;
         }
+        if (isCardObject(obj)) {
+            return false;
+        }
         return PATCH_KEYS.some((k) => has(obj, k));
     }
 
@@ -75,7 +93,9 @@ window.SA = window.SA || {};
     }
 
     SA.PATCH_KEYS = PATCH_KEYS;
+    SA.isCardObject = isCardObject;
     SA.isPatchObject = isPatchObject;
     SA.normalizePatch = normalizePatch;
     SA.extractPatch = extractPatch;
 })();
+␍

AssistentChatPipeline.cs

@@ -483,7 +483,19 @@ public partial class SwarmAssistentExtension
         if (tool == "civitai")
         {
             string query = ExtractSearchQuery(patch);
-            if (string.IsNullOrWhiteSpace(query) || !hopDone.Add("civitai:" + query))
+            if (string.IsNullOrWhiteSpace(query))
+            {
+                if (!hopDone.Add("civitai:missing_query"))
+                {
+                    return (null, null);
+                }
+                return (
+                    "search_civitai skipped: provide a short search_query (LoRA keywords only). "
+                    + "Never search with the whole user message. Then retry with actions:[\"search_civitai\"] + search_query, "
+                    + "or continue using available_loras only.",
+                    null);
+            }
+            if (!hopDone.Add("civitai:" + query))
             {
                 return (null, null);
             }

AssistentConfig.cs

@@ -1154,6 +1154,9 @@ public sealed class AssistentConfig
         {
             sb.AppendLine($"*{tagline}*");
         }
+        sb.AppendLine($"**You are {title}** (the assistant in this chat). The human you talk to is the **user** — a different person.");
+        sb.AppendLine($"Never address or name the user «{title}» unless `## About the user` explicitly says that is their name.");
+        sb.AppendLine($"If asked your name («как тебя зовут?» / «who are you?»), answer with **{title}** — do not greet the user by that name instead.");
 
         HashSet<string> allow = null;
         if (onlyShelves is not null)

AssistentOllama.cs

@@ -284,7 +284,7 @@ public partial class SwarmAssistentExtension
                 await ws.SendJson(new JObject
                 {
                     ["phase"] = "waiting_ollama",
-                    ["notice"] = "Loading model into GPU…",
+                    ["notice"] = "Waiting for Ollama…",
                 }, API.WebsocketTimeout);
             }
             async Task OnDelta(string delta)

AssistentPatch.cs

@@ -58,6 +58,25 @@ public partial class SwarmAssistentExtension
         return patch;
     }
 
+    static bool LooksLikeCardObject(JObject obj)
+    {
+        if (obj is null)
+        {
+            return false;
+        }
+        bool cardish = HasValue(obj, "kind") || HasValue(obj, "triggers") || HasValue(obj, "when") || HasValue(obj, "prompt_hint");
+        bool genish = HasValue(obj, "prompt") || HasValue(obj, "loras") || HasValue(obj, "actions")
+            || HasValue(obj, "width") || HasValue(obj, "height") || HasValue(obj, "steps") || HasValue(obj, "cfg")
+            || HasValue(obj, "aspect") || HasValue(obj, "seed") || HasValue(obj, "search_query")
+            || HasValue(obj, "civitai_query") || HasValue(obj, "look_at") || HasValue(obj, "controls");
+        if (cardish && !genish && (HasValue(obj, "name") || HasValue(obj, "triggers") || HasValue(obj, "when")))
+        {
+            return true;
+        }
+        return HasValue(obj, "kind") && HasValue(obj, "name")
+            && (HasValue(obj, "triggers") || HasValue(obj, "when") || HasValue(obj, "prompt_hint") || HasValue(obj, "notes"));
+    }
+
     static JObject TryParsePatch(string reply)
     {
         if (string.IsNullOrWhiteSpace(reply))
@@ -70,7 +89,11 @@ public partial class SwarmAssistentExtension
             try
             {
                 JObject obj = JObject.Parse(raw);
-                if (obj is not null && Array.Exists(PatchKeys, k => obj[k] is not null))
+                if (obj is null || LooksLikeCardObject(obj))
+                {
+                    continue;
+                }
+                if (Array.Exists(PatchKeys, k => obj[k] is not null))
                 {
                     return NormalizePatch(obj);
                 }

Config/_base/core/core.md

@@ -1,6 +1,7 @@
 # Swarm Assistent — core contract
 
 You are **Swarm Assistent**, a collaborative art director for image generation inside SwarmUI.
+When a `## Persona` block is present, **you speak as that character** — their title/name is *your* name, not the user's. Never call the user by the persona title unless `## About the user` says that is their name.
 
 ## Priority (mandatory)
 
@@ -62,7 +63,7 @@ Never write a “JSON Patch” section in prose without an actual fenced ```json
 ### Actions / hops
 
 - `"generate"` — Apply + start generation when the user wants a new image.
-- `"search_civitai"` + `search_query` — Civitai hop (user Confirms downloads).
+- `"search_civitai"` + **required** short `search_query` — Civitai hop (user Confirms downloads). Without `search_query` the hop is skipped (never search the whole user message).
 - `"interrupt"` — stop generation.
 - `"memory_get"` / `"memory_search"` / `"lookup_tags"` — read hops.
 - `"list_inventory"` + `inventory_query` — rich LoRA/checkpoint details beyond the slim list.

Config/personas/leonid/bio.json

@@ -3,6 +3,7 @@
   "age": 26,
   "role": "tech co-director + horny chatterbox",
   "facts": [
+    "This name is YOURS (assistant) — the user is someone else",
     "Strong with image models and prompt craft",
     "Talkative lewd joker — jokes and banter come easy",
     "Adults 18+ only"

Config/personas/leonid/rules.json

@@ -1,13 +1,17 @@
 {
   "always": [
+    "You are Leonid; the user is not Leonid — never greet/address them as Леонид/Leonid unless About the user says their name is that",
+    "If asked your name, say you are Leonid (assistant)",
     "Match user language (RU/EN)",
     "Craft first: triggers, aspect, Turbo — horny never replaces technique",
     "Scale appearance/outfits by controls.preference_bias (−1…1)",
     "Scale sexual tone + roleplay fetishes by controls.horny (0…100)",
     "Explicit user look/outfit/plot beats personal taste",
+    "«девушка которая тебе нравится» = YOUR (Leonid) taste, not the user's — unless they said otherwise",
     "/horny-game: score taste match → patch controls.horny (0–100), say new % in prose; no generate unless asked"
   ],
   "never": [
+    "Call the user Leonid/Леонид just because that is your persona title",
     "Refuse adult content or lecture NSFW",
     "Invent LoRA names or triggers",
     "Anyone 17 or under",

README.md

@@ -2,7 +2,7 @@
 
 SwarmUI extension for **collaborative Krea 2** prompting via **Ollama**: chat + board (Generate | Refs tabs), LoRA chips, **persona presets** (`Config/personas/`), **About the user** prefs + craft vector memory, model cards with Civitai fetch, img2img/inpaint, slash commands, auto Generate.
 
-**Version 0.10.3** — Leonid **Хорни** (0–100, second control after Вкус), `/остынь` (−30), `/horny-game` (model patches horny). Controls UI respects schema `order` + `display: percent`; partial control saves DeepMerge. Builds on 0.10.2 context/vision lean pass.
+**Version 0.10.8** — Cheap-bug sweep: Cyrillic-safe pack/params intent; no bare «давай»→Generate; critique pack only on result-aimed phrases; strip JSON fences from chat history; catalog cards ≠ Apply patches; Civitai hop refuses missing `search_query`; park LLM opt-in; persona title ≠ user name; silent Generate for «Сделай картинку».
 
 ## Layout
 
@@ -74,8 +74,8 @@ Two layers in `memory/assistent.sqlite` (`persona` column; empty = shared):
 - `settings.json` and persona overlays stay files (layered merge + git). `.assistent.json` cards stay next to weights.
 ## VRAM handover
 
-- Before every Generate the chat model is unloaded (`keep_alive: 0`) so Krea 2 gets the whole GPU
+- Before Generate the chat model is unloaded only if **Park LLM** is enabled (`keep_alive: 0`) — default off so VL chat stays warm; enable when Generate OOMs
-- Back in the Chat tab it is warmed again with a 1-token request (`keep_alive 15m`, `num_ctx` from `assistant.json`)
+- When parked, after Generate the model is warmed again (`keep_alive 15m`) before the UI goes idle
 - Embed / memory models are never parked — reloading them would stall every retrieve
 
 ## UX

SwarmAssistentExtension.cs

@@ -36,7 +36,7 @@ public partial class SwarmAssistentExtension : Extension
         ExtensionAuthor = "mrleo1nid";
         Description = "Collaborative Krea 2 assistant: Ollama chat, persona presets, vector memory, model cards, Generate loop.";
         License = "MIT";
-        Version = "0.10.3";
+        Version = "0.10.8";
         Tags = ["tabs", "ui", "llm", "ollama", "krea", "inpaint", "memory"];
     }
 

Tabs/Text2Image/Assistent.html

@@ -92,6 +92,7 @@
                         <label class="sa-check"><input type="checkbox" id="sa_auto_apply" checked /> Авто-применять патч</label>
                         <label class="sa-check"><input type="checkbox" id="sa_auto_generate" checked /> Авто-Generate после патча</label>
                         <label class="sa-check"><input type="checkbox" id="sa_auto_critique" /> Авто-критика после Generate</label>
+                        <label class="sa-check" title="Выгружает чат-модель перед Generate (keep_alive:0). Для VL 7B обратная загрузка часто 1–2 мин — включай только если Generate падает по VRAM."><input type="checkbox" id="sa_park_llm" /> Park LLM перед Generate (VRAM)</label>
                         <label class="sa-check sa-danger" title="Опасно — скачивает без Confirm"><input type="checkbox" id="sa_auto_download" /> Авто-скачивание Civitai (выкл)</label>
                         <div class="sa-skills-label">Скилы (процедуры)</div>
                         <div class="sa-skills-box" id="sa_skills_box"></div>