feat(models): auto-mode intent remap + real coder models in preference

Settings "Auto model routing" picks which installed model fires for chat vs
coding intent when no model is pinned (auto_chat_model / auto_code_model).
_auto_select_model honors the remap; _MODEL_PREFERENCE["code"] prefers real
coder models first.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

interface/web/src/Settings.jsx

@@ -3,6 +3,8 @@ import { API_BASE } from "./config";
 
 const DEFAULTS = {
   model: "",
+  auto_chat_model: "", // Auto-mode: model for chat intent ("" = built-in preference)
+  auto_code_model: "", // Auto-mode: model for code intent
   think: false, // Qwen3-style reasoning; off = much faster chat/memory
   temperature: 0.7,
   num_ctx: 0, // context window in tokens; 0 = model default
@@ -30,8 +32,11 @@ export function Settings() {
   const [selectedApps, setSelectedApps] = useState([]);
   const [brandQueue, setBrandQueue] = useState([]);
   const [isBranding, setIsBranding] = useState(false);
+  const [modelList, setModelList] = useState([]);
 
   useEffect(() => {
+    fetch(`${API_BASE}/models`).then(r => r.ok ? r.json() : null)
+      .then(d => setModelList((d && d.models) || [])).catch(() => {});
     fetch(`${API_BASE}/settings`)
       .then(r => r.ok ? r.json() : null)
       .then(d => { if (d) setForm(f => ({ ...f, ...d })); })
@@ -165,6 +170,29 @@ export function Settings() {
     <div style={{ maxWidth: "720px" }}>
       <h2 style={{ margin: "0 0 1.25rem", fontSize: "1.1rem", color: "#eee" }}>Settings</h2>
 
+      {/* Auto model routing */}
+      <div style={sectionStyle}>
+        <h3 style={{ margin: "0 0 0.5rem", fontSize: "0.95rem", color: "#bbb" }}>Auto model routing</h3>
+        <div style={{ fontSize: "0.72rem", color: "#555", marginBottom: "0.75rem" }}>
+          When no model is pinned (chat picker on "auto"), which model fires for each
+          detected intent. "Auto" = the built-in preference for your hardware.
+          {form.model && <span style={{ color: "#c9a227" }}> A model is currently pinned in chat, so routing is bypassed until you set it back to auto.</span>}
+        </div>
+        <div style={{ display: "flex", gap: "1rem" }}>
+          {[["auto_chat_model", "Chat / general"], ["auto_code_model", "Coding questions"]].map(([key, label]) => (
+            <div key={key} style={{ flex: 1 }}>
+              <label style={labelStyle}>{label}</label>
+              <select value={form[key]} onChange={e => update(key, e.target.value)}
+                style={{ width: "100%", padding: "0.6rem", background: "#222", color: "#eee", border: "1px solid #333", borderRadius: "8px", marginTop: "0.3rem" }}>
+                <option value="">Auto (best installed)</option>
+                {modelList.map(m => <option key={m} value={m}>{m}</option>)}
+                {form[key] && !modelList.includes(form[key]) && <option value={form[key]}>{form[key]} (not installed)</option>}
+              </select>
+            </div>
+          ))}
+        </div>
+      </div>
+
       {/* Generation */}
       <div style={sectionStyle}>
         <h3 style={{ margin: "0 0 1rem", fontSize: "0.95rem", color: "#bbb" }}>Generation</h3>

synapse/main.py

@@ -104,6 +104,11 @@ async def _auto_select_model(message: str = "") -> str:
         if s.get("model"):
             return s["model"]
         intent = _detect_intent(message) if message else "chat"
+        # Auto-mode remap: a configured model for this intent fires first;
+        # otherwise fall back to the built-in preference list.
+        remap = s.get(f"auto_{intent}_model")
+        if remap:
+            return remap
         return await get_ollama_manager().select_best_model(intent)
     except Exception:
         return DEFAULT_CHAT_MODEL

synapse/memory/store.py

@@ -986,6 +986,10 @@ class PersistentMemoryStore:
     # -----------------------------
     _SETTINGS_DEFAULTS: Dict[str, Any] = {
         "model": "",
+        # Auto-mode routing overrides (blank = built-in preference). When no model
+        # is pinned, a message's detected intent picks which of these fires first.
+        "auto_chat_model": "",
+        "auto_code_model": "",
         # Qwen3-style reasoning. Off by default: the hidden <think> block is pure
         # latency for chat/memory. Turn on for hard multi-step problems.
         "think": False,

synapse/ollama_manager.py

@@ -162,7 +162,7 @@ def _detect_gpu_backend() -> tuple[str, dict]:
 # tail differs by task. Prefix-matched against installed model names.
 _MODEL_PREFERENCE = {
     "chat": ("qwen2.5:3b", "qwen2.5", "gemma3:1b", "gemma3", "phi3", "phi-3", "gemma2", "gemma"),
-    "code": ("qwen2.5:3b", "qwen2.5", "gemma3:1b", "gemma3", "phi3", "phi-3", "codellama", "deepseek-coder", "codegemma"),
+    "code": ("qwen2.5-coder", "qwen3-coder", "deepseek-coder", "codellama", "codegemma", "qwen2.5:3b", "qwen2.5", "gemma3", "phi3", "phi-3"),
 }
 
 

tests/test_smoke.py

@@ -43,6 +43,19 @@ def test_keep_alive_pins_the_model():
     assert "keep_alive" not in mgr._apply_keep_alive({"model": "x"})
 
 
+def test_auto_model_remap(monkeypatch):
+    import asyncio
+    from synapse import main
+    cfg = {"model": "", "auto_chat_model": "chatX", "auto_code_model": "coderY"}
+    monkeypatch.setattr(main.store, "get_settings", lambda: cfg)
+    # code intent ("function") routes to the code remap; chat intent to the chat remap
+    assert asyncio.run(main._auto_select_model("write a function to sort a list")) == "coderY"
+    assert asyncio.run(main._auto_select_model("how are you today")) == "chatX"
+    # an explicit pin beats the remap
+    cfg["model"] = "pinnedZ"
+    assert asyncio.run(main._auto_select_model("debug this code")) == "pinnedZ"
+
+
 def test_hardware_fit_logic():
     from synapse import hardware
     assert hardware._fit(2.5, 4.0, 16.0) == "gpu"    # 2.5+1 <= 4 -> fits GPU