Ship Assistent 0.14.1: Exact generate params and cheap park/warm.

Client always merges Exact turbo|raw steps/cfg/sigma before Generate so sparse LLM omissions and leftover SD 20/7 cannot stick; Ollama park/warm skip no-op /api/ps round-trips when the chat model is already unloaded or resident.

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

AssistentVram.cs

@@ -10,7 +10,9 @@ namespace Mrleo1nid.SwarmAssistent;
 
 /// <summary>VRAM handover between Ollama and the image backend: park the chat model before
 /// Generate, warm it again once the user is back in the chat. Embed / memory models are never
-/// parked — they are tiny and reloading them stalls every retrieve.</summary>
+/// parked — they are tiny and reloading them stalls every retrieve.
+/// Warm is a no-op when Ollama already has the chat model resident (/api/ps) so post-Generate
+/// reload is cheap when Krea did not evict VL.</summary>
 public partial class SwarmAssistentExtension
 {
     const string WarmKeepAlive = "15m";
@@ -28,6 +30,18 @@ public partial class SwarmAssistentExtension
         {
             return new JObject { ["success"] = true, ["parked"] = false, ["skipped"] = "memory model — never parked" };
         }
+        // Already unloaded — skip the keep_alive:0 round-trip.
+        if (!await IsOllamaModelResident(root, name))
+        {
+            return new JObject
+            {
+                ["success"] = true,
+                ["parked"] = false,
+                ["skipped"] = "not_resident",
+                ["model"] = name,
+                ["base_url"] = root,
+            };
+        }
         JObject generate = new()
         {
             ["model"] = name,
@@ -56,7 +70,8 @@ public partial class SwarmAssistentExtension
         return new JObject { ["success"] = true, ["parked"] = true, ["model"] = name, ["base_url"] = root };
     }
 
-    /// <summary>Single-token chat so the model is resident again by the time the user types.</summary>
+    /// <summary>Single-token chat so the model is resident again by the time the user types.
+    /// Skips the load when /api/ps already lists the model (avoids ~20–30s no-op warm).</summary>
     public async Task<JObject> AssistentWarmLlm(Session session, string baseUrl, string model)
     {
         string root = NormalizeBaseUrl(baseUrl);
@@ -69,6 +84,17 @@ public partial class SwarmAssistentExtension
         {
             return new JObject { ["success"] = true, ["warmed"] = false, ["skipped"] = "memory model" };
         }
+        if (await IsOllamaModelResident(root, name))
+        {
+            return new JObject
+            {
+                ["success"] = true,
+                ["warmed"] = true,
+                ["skipped"] = "already_resident",
+                ["model"] = name,
+                ["keep_alive"] = WarmKeepAlive,
+            };
+        }
         int numCtx = CfgInt("num_ctx", DefaultNumCtxFallback);
         JObject payload = new()
         {
@@ -101,6 +127,50 @@ public partial class SwarmAssistentExtension
         };
     }
 
+    /// <summary>True when Ollama /api/ps lists <paramref name="model"/> (or a matching tag).</summary>
+    async Task<bool> IsOllamaModelResident(string root, string model)
+    {
+        if (string.IsNullOrWhiteSpace(root) || string.IsNullOrWhiteSpace(model))
+        {
+            return false;
+        }
+        try
+        {
+            using HttpResponseMessage resp = await HttpClient.GetAsync($"{root}/api/ps");
+            if (!resp.IsSuccessStatusCode)
+            {
+                return false;
+            }
+            string body = await resp.Content.ReadAsStringAsync();
+            JObject parsed = JObject.Parse(body);
+            JArray models = parsed["models"] as JArray ?? [];
+            foreach (JToken m in models)
+            {
+                string name = m["name"]?.ToString() ?? m["model"]?.ToString() ?? "";
+                if (string.IsNullOrWhiteSpace(name))
+                {
+                    continue;
+                }
+                if (string.Equals(name, model, StringComparison.OrdinalIgnoreCase))
+                {
+                    return true;
+                }
+                // Tags may differ by :latest vs bare name.
+                if (name.StartsWith(model + ":", StringComparison.OrdinalIgnoreCase)
+                    || model.StartsWith(name + ":", StringComparison.OrdinalIgnoreCase))
+                {
+                    return true;
+                }
+            }
+            return false;
+        }
+        catch (Exception ex)
+        {
+            Logs.Debug($"IsOllamaModelResident: {ex.Message}");
+            return false;
+        }
+    }
+
     static async Task<(bool ok, string body)> PostOllamaJson(string root, string route, JObject payload)
     {
         try
@@ -108,7 +178,7 @@ public partial class SwarmAssistentExtension
             using StringContent content = new(payload.ToString(Newtonsoft.Json.Formatting.None), Encoding.UTF8, "application/json");
             using HttpResponseMessage resp = await HttpClient.PostAsync($"{root}{route}", content);
             string body = await resp.Content.ReadAsStringAsync();
-            return (resp.IsSuccessStatusCode, resp.IsSuccessStatusCode ? body : $"HTTP {(int)resp.StatusCode}: {body}");
+            return (resp.IsSuccessStatusCode, body);
         }
         catch (Exception ex)
         {