Harden Exact reconcile and soften post-Generate warm.

Overwrite foreign steps/cfg in sparse patches when the user did not ask for params, and skip the cold-load path when Ollama already has the chat model resident.

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

src/app.js

@@ -1353,7 +1353,7 @@
                 if (want == null) {
                     return false;
                 }
-                if (patch?.[k] != null && String(patch[k]) !== String(want)) {
+                if (userIntent && patch?.[k] != null && String(patch[k]) !== String(want)) {
                     return false;
                 }
                 if (userIntent && state.sessionExact?.[k] != null) {
@@ -4751,20 +4751,25 @@
     }
 
     /** Re-load chat model into VRAM. force=true after Generate even without park — Krea often evicts Ollama. */
+    /**
+     * Re-load chat model into VRAM. force=true after Generate even without park —
+     * Krea often evicts Ollama. Soft: AssistentWarmLlm skipped=already_resident
+     * clears expectColdLoad without a fake cold path. Resolves { ok, alreadyResident }.
+     */
     function warmLlm({ force = false } = {}) {
         return new Promise((resolve) => {
             const model = $('sa_model')?.value;
             if (!model || typeof genericRequest !== 'function') {
-                resolve(false);
+                resolve({ ok: false, alreadyResident: false });
                 return;
             }
             if (!force && !state.llmParked) {
-                resolve(false);
+                resolve({ ok: false, alreadyResident: false });
                 return;
             }
             const baseUrl = $('sa_base_url')?.value || 'http://127.0.0.1:11434';
             let settled = false;
-            const finish = (ok) => {
+            const finish = (ok, alreadyResident = false) => {
                 if (settled) {
                     return;
                 }
@@ -4773,11 +4778,20 @@
                 if (ok) {
                     state.expectColdLoad = false;
                 }
-                resolve(!!ok);
+                resolve({ ok: !!ok, alreadyResident: !!alreadyResident });
             };
             // VL 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));
+            genericRequest(
+                'AssistentWarmLlm',
+                { baseUrl, model },
+                (data) => {
+                    const skipped = !!(data && (data.skipped === 'already_resident' || data.already_resident === true));
+                    finish(true, skipped);
+                },
+                0,
+                () => finish(false),
+            );
         });
     }
 
@@ -5130,7 +5144,10 @@
 
         state.generating = false;
         setInterruptVisible(state.busy);
-        state.expectColdLoad = true;
+        // Soft cold flag: only assume eviction when we parked. Otherwise wait for
+        // warmLlm — already_resident means keep chatting without a cold path.
+        const parkedBeforeWarm = !!state.llmParked;
+        state.expectColdLoad = parkedBeforeWarm;
 
         if (jobs && lastSrc && state.selectedGenResultId && epoch === state.chatEpoch) {
             const row = state.genResults.find((r) => r.id === state.selectedGenResultId);
@@ -5144,9 +5161,16 @@
         const multiDone = !!(jobs && finishedGenResultCount() > 1);
         const willAutoCritique = !multiDone && !!$('sa_auto_critique')?.checked;
         if (state.view === 'chat' && paneVisible && !willAutoCritique && epoch === state.chatEpoch) {
-            startBusyUi('warming');
-            setStatus('Возвращаю LLM в GPU…');
-            await warmLlm({ force: true });
+            if (parkedBeforeWarm || state.expectColdLoad) {
+                startBusyUi('warming');
+                setStatus('Возвращаю LLM в GPU…');
+            }
+            const warmResult = await warmLlm({ force: true });
+            if (warmResult?.alreadyResident) {
+                state.expectColdLoad = false;
+            } else if (!warmResult?.ok) {
+                state.expectColdLoad = true;
+            }
         }
 
         if (epoch !== state.chatEpoch) {

src/kreaProfile.js

@@ -81,10 +81,16 @@ export function exactKeysToForce(live, defaults, {
     if (want == null) {
       continue;
     }
-    const eps = key === 'steps' ? 0.5 : 0.051;
-    // Patch asked for a non-Exact number — honor it.
-    if (patch && patch[key] != null && !numClose(patch[key], want, eps)) {
-      continue;
+    const eps = key === 'steps' ? 0.5 : 0.051;
+    // Honor intentional non-Exact patch only when the user asked for params.
+    // Sparse LLM echoes of live 20/7 must not block Exact (turbo/raw label -> numbers).
+    if (
+      userParamIntent
+      && patch
+      && patch[key] != null
+      && !numClose(patch[key], want, eps)
+    ) {
+      continue;
     }
     if (userParamIntent && sessionExact && sessionExact[key] != null) {
       continue;

src/session.js

@@ -372,7 +372,7 @@ export function resolveExactProfileDefaults({ exact, profiles, profileName } = {
 }
 
 /**
- * On generate, inject Exact profile steps/cfg/sigma when the patch omitted them.
+ * On generate, inject Exact profile steps/cfg/sigma when omitted, and overwrite foreign leftovers when the user did not ask for params.
  * Client is authoritative so sparse LLM deltas are safe. Honors explicit patch values
  * and sessionExact when the user asked for params this turn.
  * Returns clearSessionKeys so the UI apply path is not blocked by stale sessionExact.
@@ -390,9 +390,20 @@ export function mergeExactParamsForGenerate(patch, {
   const defaults = resolveExactProfileDefaults({ exact, profiles, profileName });
   const out = { ...patch };
   const clearSessionKeys = [];
-  for (const key of EXACT_GENERATE_PARAM_KEYS) {
-    if (out[key] != null) {
-      continue;
+  for (const key of EXACT_GENERATE_PARAM_KEYS) {
+    if (out[key] != null) {
+      // Sparse LLM may echo live leftovers (20/7). Without user intent, force Exact.
+      if (
+        !userParamIntent
+        && defaults[key] != null
+        && String(out[key]) !== String(defaults[key])
+      ) {
+        out[key] = defaults[key];
+        if (sessionExact?.[key] != null && String(sessionExact[key]) !== String(defaults[key])) {
+          clearSessionKeys.push(key);
+        }
+      }
+      continue;
     }
     if (userParamIntent && sessionExact?.[key] != null) {
       out[key] = sessionExact[key];