fix: duplicate assistant replies and flickering conversation list

backend/src/agent/builder.py

@@ -28,8 +28,19 @@ class AgentDeps:
     os.environ.setdefault("GEMINI_API_KEY", Settings.GEMINI_API_KEY)
     # `google-gla:` was dropped in pydantic-ai 2.x — the prefix is now `google:`
     model_name = Settings.GEMINI_MODEL_NAME or "google:gemini-3-flash-preview"
+    # Reasoning tokens share the output budget on Gemini's thinking models.
+    # Capped below max_tokens so the post-tool-call turn always has room left
+    # to answer — uncapped, the model has burned the whole budget thinking,
+    # returned neither text nor a tool call, and pydantic-ai's forced retry
+    # made it re-answer from scratch: duplicated reply, then a hard failure
+    # once retries ran out ("Exceeded maximum output retries").
+    _reasoning_settings = {"google_thinking_config": {"thinking_budget": 8192}}
 else:
     model_name = Settings.OPENAI_MODEL_NAME or "openai:gpt-5.2"
+    # Same failure class applies to OpenAI's reasoning models (o-series/gpt-5):
+    # reasoning tokens count against max_completion_tokens too. No numeric
+    # budget knob here, only an effort level — capped so it can't dominate.
+    _reasoning_settings = {"openai_reasoning_effort": "low"}
 
 _agent_cache: dict[str, Agent] = {}
 
@@ -40,11 +51,11 @@ def build_agent(city) -> Agent:
         model_name,
         system_prompt=city.system_prompt + UI_TOOLS_PROMPT,
         deps_type=AgentDeps,
-        # Without an explicit cap the provider default applies, and on Gemini's
-        # thinking models reasoning tokens come out of that same output budget —
-        # the model can spend the whole allowance thinking and die with
-        # "token limit exceeded before any response was generated".
-        model_settings=ModelSettings(max_tokens=8192),
+        # Without an explicit cap the provider default applies — see
+        # `_reasoning_settings` above for why that's dangerous on a thinking
+        # model. 16384 gives that capped reasoning budget headroom to still
+        # leave room for the actual reply.
+        model_settings=ModelSettings(max_tokens=16384, **_reasoning_settings),
     )
 
     search.register(agent, city)

frontend/app/components/ConversationSidebar.tsx

@@ -153,8 +153,11 @@ export default function ConversationSidebar({
     router.replace("/login");
   }
 
-  const fetchPage = useCallback(async (pageNum: number, append = false) => {
-    if (pageNum === 1) { setLoading(true); setError(null); }
+  const fetchPage = useCallback(async (pageNum: number, append = false, silent = false) => {
+    // `silent` is for background refreshes (e.g. after switching threads) where
+    // we already have data on screen — showing the skeleton would blank a list
+    // the user is actively looking at for no reason.
+    if (pageNum === 1 && !silent) { setLoading(true); setError(null); }
     try {
       const res = await fetch(`/api/conversations?limit=10&page=${pageNum}`, { credentials: "include" });
       if (!res.ok) throw new Error(`${res.status}`);
@@ -163,10 +166,10 @@ export default function ConversationSidebar({
       setConversations(prev => append ? [...prev, ...convs] : convs);
       setHasMore(Array.isArray(data) ? false : (data.has_more ?? false));
     } catch (e) {
-      setError("Could not load conversations.");
+      if (!silent) setError("Could not load conversations.");
       console.error("[ConversationSidebar]", e);
     } finally {
-      if (pageNum === 1) setLoading(false);
+      if (pageNum === 1 && !silent) setLoading(false);
     }
   }, []);
 
@@ -204,8 +207,12 @@ export default function ConversationSidebar({
 
   useEffect(() => { fetchPage(1); }, [fetchPage]);
 
+  // Refresh the list after switching threads (title/count may have changed) —
+  // skip the very first run, the mount effect above already just fetched it.
+  const didMount = useRef(false);
   useEffect(() => {
-    const t = setTimeout(() => fetchPage(1), 1500);
+    if (!didMount.current) { didMount.current = true; return; }
+    const t = setTimeout(() => fetchPage(1, false, true), 1500);
     return () => clearTimeout(t);
   }, [activeThreadId, fetchPage]);