Fix Ollama chat reliability

backend/app/agent.py

@@ -24,6 +24,12 @@ from .tools import TOOL_DEFINITIONS, ToolError, run_tool
 MAX_ITERATIONS = 6
 
 
+def _tools_not_supported(exc: OllamaError) -> bool:
+    """Détecte l'erreur Ollama renvoyée par un modèle sans function calling."""
+    message = str(exc).lower()
+    return "does not support tools" in message or "does not support tool" in message
+
+
 def _parse_args(raw) -> dict:
     if isinstance(raw, dict):
         return raw
@@ -54,26 +60,62 @@ async def run_agent(
         tools = None
     collected: list[dict] = []
     text_parts: list[str] = []
+    active_tools = tools
+    tool_fallback_used = False
 
     try:
         for _ in range(MAX_ITERATIONS):
             content_buf = ""
+            thinking_buf = ""
+            thinking_status_sent = False
             tool_calls: list[dict] = []
 
-            async for chunk in ollama.chat(
-                model, convo, tools=tools, options=options, stream=True
-            ):
-                msg = chunk.get("message", {})
-                token = msg.get("content", "")
-                if token:
-                    content_buf += token
-                    yield {"type": "token", "content": token}
-                if msg.get("tool_calls"):
-                    tool_calls.extend(msg["tool_calls"])
-                if chunk.get("done"):
+            # Un modèle peut savoir discuter sans supporter les outils. Ollama
+            # refuse alors la requête entière : on retente une fois en chat simple.
+            while True:
+                try:
+                    async for chunk in ollama.chat(
+                        model, convo, tools=active_tools, options=options, stream=True
+                    ):
+                        msg = chunk.get("message", {})
+                        token = msg.get("content", "")
+                        thinking = msg.get("thinking", "")
+                        if thinking:
+                            thinking_buf += thinking
+                            if not thinking_status_sent:
+                                thinking_status_sent = True
+                                yield {"type": "status", "message": "Réflexion…"}
+                        if token:
+                            content_buf += token
+                            yield {"type": "token", "content": token}
+                        if msg.get("tool_calls"):
+                            tool_calls.extend(msg["tool_calls"])
+                        if chunk.get("done"):
+                            break
                     break
+                except OllamaError as exc:
+                    if (
+                        active_tools
+                        and not tool_fallback_used
+                        and not content_buf
+                        and not tool_calls
+                        and _tools_not_supported(exc)
+                    ):
+                        active_tools = None
+                        tool_fallback_used = True
+                        yield {
+                            "type": "notice",
+                            "message": (
+                                "Ce modèle ne supporte pas les outils ; "
+                                "réponse en mode conversation simple."
+                            ),
+                        }
+                        continue
+                    raise
 
             assistant_turn: dict = {"role": "assistant", "content": content_buf}
+            if thinking_buf:
+                assistant_turn["thinking"] = thinking_buf
             if tool_calls:
                 assistant_turn["tool_calls"] = tool_calls
             convo.append(assistant_turn)
@@ -106,7 +148,7 @@ async def run_agent(
                     convo.append(
                         {
                             "role": "tool",
-                            "name": name,
+                            "tool_name": name,
                             "content": json.dumps(
                                 {
                                     "ok": False,
@@ -137,7 +179,7 @@ async def run_agent(
                 convo.append(
                     {
                         "role": "tool",
-                        "name": name,
+                        "tool_name": name,
                         "content": json.dumps(result, ensure_ascii=False),
                     }
                 )
@@ -162,11 +204,21 @@ async def run_agent(
         yield {"type": "error", "message": f"Ollama : {exc}"}
         return
     except (httpx.HTTPError, OSError) as exc:
-        yield {"type": "error", "message": str(exc)}
+        yield {
+            "type": "error",
+            "message": f"Impossible de joindre Ollama ({ollama.host}) : {exc}",
+        }
         return
 
-    yield {
-        "type": "final",
-        "content": "\n\n".join(text_parts).strip(),
-        "tools": collected,
-    }
+    final_content = "\n\n".join(text_parts).strip()
+    if not final_content and not collected:
+        yield {
+            "type": "error",
+            "message": (
+                "Le modèle a terminé sans renvoyer de texte. Essayez un modèle "
+                "de chat récent ou désactivez son mode de réflexion avancée."
+            ),
+        }
+        return
+
+    yield {"type": "final", "content": final_content, "tools": collected}

backend/app/routes/chat.py

@@ -9,6 +9,7 @@ Flux :
 from __future__ import annotations
 
 import json
+import logging
 
 from fastapi import APIRouter, HTTPException
 from fastapi.responses import StreamingResponse
@@ -19,6 +20,7 @@ from ..agent import run_agent
 from ..config import settings
 
 router = APIRouter(prefix="/api", tags=["chat"])
+logger = logging.getLogger(__name__)
 
 
 class ChatRequest(BaseModel):
@@ -54,20 +56,38 @@ async def chat(req: ChatRequest) -> StreamingResponse:
         yield _sse("start", {"model": model})
         final_content = ""
         tools_meta: list[dict] = []
+        error_message = ""
 
-        async for ev in run_agent(
-            model,
-            convo,
-            options=agent_config.ollama_options(cfg),
-            enabled_tools=agent_config.enabled_tool_names(cfg),
-            confirm_shell=cfg.get("confirm_shell", True),
-        ):
-            etype = ev.pop("type")
-            if etype in ("token", "tool_call", "tool_result", "tool_confirm", "error"):
-                yield _sse(etype, ev)
-            elif etype == "final":
-                final_content = ev["content"]
-                tools_meta = ev["tools"]
+        try:
+            async for ev in run_agent(
+                model,
+                convo,
+                options=agent_config.ollama_options(cfg),
+                enabled_tools=agent_config.enabled_tool_names(cfg),
+                confirm_shell=cfg.get("confirm_shell", True),
+            ):
+                etype = ev.pop("type")
+                if etype in (
+                    "token",
+                    "status",
+                    "notice",
+                    "tool_call",
+                    "tool_result",
+                    "tool_confirm",
+                ):
+                    yield _sse(etype, ev)
+                elif etype == "error":
+                    error_message = ev.get("message", "Erreur Ollama inconnue")
+                    yield _sse("error", {"message": error_message})
+                elif etype == "final":
+                    final_content = ev["content"]
+                    tools_meta = ev["tools"]
+        except Exception as exc:
+            # Après l'envoi des en-têtes SSE, une exception non gérée coupe le
+            # socket sans explication et laisse l'interface bloquée.
+            logger.exception("Échec inattendu du flux de chat")
+            error_message = f"Erreur interne du chat : {exc}"
+            yield _sse("error", {"message": error_message})
 
         if final_content or tools_meta:
             db.add_message(
@@ -77,6 +97,20 @@ async def chat(req: ChatRequest) -> StreamingResponse:
                 model,
                 meta={"tools": tools_meta} if tools_meta else None,
             )
-        yield _sse("done", {"content": final_content, "tools": tools_meta})
+        yield _sse(
+            "done",
+            {
+                "content": final_content,
+                "tools": tools_meta,
+                "error": error_message or None,
+            },
+        )
 
-    return StreamingResponse(event_stream(), media_type="text/event-stream")
+    return StreamingResponse(
+        event_stream(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache, no-transform",
+            "X-Accel-Buffering": "no",
+        },
+    )

frontend/src/api/client.ts

@@ -176,15 +176,36 @@ export async function streamChat(
     onToolCall: (call: ToolCall) => void;
     onToolResult: (call: ToolCall) => void;
     onToolConfirm: (command: string) => void;
+    onStatus: (msg: string) => void;
+    onNotice: (msg: string) => void;
     onDone: (full: string) => void;
     onError: (msg: string) => void;
   }
 ): Promise<void> {
-  const res = await fetch("/api/chat", {
-    method: "POST",
-    headers: { "Content-Type": "application/json" },
-    body: JSON.stringify(body),
-  });
+  let res: Response;
+  try {
+    res = await fetch("/api/chat", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    });
+  } catch (err) {
+    handlers.onError(
+      err instanceof Error ? err.message : "serveur Loki injoignable"
+    );
+    return;
+  }
+  if (!res.ok) {
+    let message = `requête refusée (${res.status})`;
+    try {
+      const payload = await res.json();
+      message = payload.detail ?? payload.error ?? message;
+    } catch {
+      /* réponse non JSON */
+    }
+    handlers.onError(message);
+    return;
+  }
   if (!res.body) {
     handlers.onError("pas de flux de réponse");
     return;
@@ -193,35 +214,71 @@ export async function streamChat(
   const reader = res.body.getReader();
   const decoder = new TextDecoder();
   let buffer = "";
+  let terminal = false;
+  let failed = false;
 
-  while (true) {
-    const { done, value } = await reader.read();
-    if (done) break;
-    buffer += decoder.decode(value, { stream: true });
-    const events = buffer.split("\n\n");
-    buffer = events.pop() ?? "";
-
-    for (const block of events) {
-      let event = "message";
-      let data = "";
-      for (const line of block.split("\n")) {
-        if (line.startsWith("event: ")) event = line.slice(7).trim();
-        else if (line.startsWith("data: ")) data += line.slice(6);
+  const dispatch = (raw: string) => {
+    const block = raw.replace(/\r\n/g, "\n");
+    let event = "message";
+    const dataLines: string[] = [];
+    for (const line of block.split("\n")) {
+      if (line.startsWith("event:")) event = line.slice(6).trim();
+      else if (line.startsWith("data:")) dataLines.push(line.slice(5).trimStart());
+    }
+    if (dataLines.length === 0) return;
+    try {
+      const payload = JSON.parse(dataLines.join("\n"));
+      if (event === "token") handlers.onToken(payload.content);
+      else if (event === "status") handlers.onStatus(payload.message);
+      else if (event === "notice") handlers.onNotice(payload.message);
+      else if (event === "tool_call")
+        handlers.onToolCall({ ...payload, status: "running" });
+      else if (event === "tool_result") handlers.onToolResult(payload);
+      else if (event === "tool_confirm") handlers.onToolConfirm(payload.command);
+      else if (event === "error") {
+        failed = true;
+        handlers.onError(payload.message);
+      } else if (event === "done") {
+        terminal = true;
+        if (payload.error) {
+          if (!failed) handlers.onError(payload.error);
+          failed = true;
+        } else if (!failed) {
+          handlers.onDone(payload.content);
+        }
       }
-      if (!data) continue;
-      try {
-        const payload = JSON.parse(data);
-        if (event === "token") handlers.onToken(payload.content);
-        else if (event === "tool_call")
-          handlers.onToolCall({ ...payload, status: "running" });
-        else if (event === "tool_result") handlers.onToolResult(payload);
-        else if (event === "tool_confirm") handlers.onToolConfirm(payload.command);
-        else if (event === "done") handlers.onDone(payload.content);
-        else if (event === "error") handlers.onError(payload.message);
-      } catch {
-        /* bloc partiel */
+    } catch {
+      if (!failed) {
+        failed = true;
+        handlers.onError("réponse illisible reçue du serveur");
       }
     }
+  };
+
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      buffer += decoder.decode(value, { stream: true }).replace(/\r\n/g, "\n");
+      const events = buffer.split("\n\n");
+      buffer = events.pop() ?? "";
+      for (const block of events) {
+        if (block.trim()) dispatch(block);
+      }
+    }
+    buffer += decoder.decode();
+    if (buffer.trim()) dispatch(buffer);
+  } catch (err) {
+    if (!failed) {
+      failed = true;
+      handlers.onError(
+        err instanceof Error ? err.message : "connexion au chat interrompue"
+      );
+    }
+  }
+
+  if (!terminal && !failed) {
+    handlers.onError("le serveur a fermé la réponse avant sa fin");
   }
 }
 

frontend/src/panels/ChatPanel.tsx

@@ -11,6 +11,8 @@ export function ChatPanel() {
     messages,
     streaming,
     streamContent,
+    streamStatus,
+    streamNotice,
     streamTools,
     sendMessage,
     currentSessionId,
@@ -101,6 +103,8 @@ export function ChatPanel() {
                   created_at: Date.now() / 1000,
                 }}
                 pending
+                pendingStatus={streamStatus}
+                notice={streamNotice}
               />
             )}
             {pendingShell && (
@@ -154,7 +158,17 @@ export function ChatPanel() {
   );
 }
 
-function Bubble({ msg, pending }: { msg: Message; pending?: boolean }) {
+function Bubble({
+  msg,
+  pending,
+  pendingStatus,
+  notice,
+}: {
+  msg: Message;
+  pending?: boolean;
+  pendingStatus?: string;
+  notice?: string | null;
+}) {
   const time = new Date(msg.created_at * 1000).toLocaleTimeString("fr-FR", {
     hour: "2-digit",
     minute: "2-digit",
@@ -198,9 +212,16 @@ function Bubble({ msg, pending }: { msg: Message; pending?: boolean }) {
         {(msg.meta?.tools ?? []).map((t: ToolCall, i: number) => (
           <ToolCard key={i} call={t} />
         ))}
+        {notice && (
+          <div className="mb-2 rounded-lg border border-[#4a3a2a] bg-card-deep px-3 py-2 text-xs text-warn">
+            {notice}
+          </div>
+        )}
         {(msg.content || pending) && (
           <div className="text-sm leading-[1.65] text-ink-2 whitespace-pre-wrap">
-            {msg.content}
+            {msg.content || (
+              <span className="text-muted-2">{pendingStatus || "Génération…"}</span>
+            )}
             {pending && (
               <span className="ml-0.5 inline-block h-3.5 w-[7px] animate-pulse bg-accent align-middle" />
             )}

frontend/src/store/useStore.ts

@@ -34,6 +34,8 @@ interface LokiState {
   messages: Message[];
   streaming: boolean;
   streamContent: string; // réponse de l'agent en cours de frappe
+  streamStatus: string;
+  streamNotice: string | null;
   streamTools: ToolCall[]; // appels d'outils de la réponse en cours
 
   fileTree: FileNode[];
@@ -77,6 +79,8 @@ export const useStore = create<LokiState>((set, get) => ({
   messages: [],
   streaming: false,
   streamContent: "",
+  streamStatus: "",
+  streamNotice: null,
   streamTools: [],
   fileTree: [],
   previewPath: null,
@@ -154,9 +158,6 @@ export const useStore = create<LokiState>((set, get) => ({
     try {
       const status = await getStatus();
       set({ status });
-      if (!get().selectedModel && status.default_model) {
-        set({ selectedModel: status.default_model });
-      }
     } catch {
       set({ status: { connected: false, host: "", default_model: "" } });
     }
@@ -166,8 +167,14 @@ export const useStore = create<LokiState>((set, get) => ({
     set({ loadingModels: true });
     try {
       const { models, default: def } = await getModels();
-      set({ models });
-      if (!get().selectedModel && def) set({ selectedModel: def });
+      const installed = new Set(models.map((model) => model.name));
+      const current = get().selectedModel;
+      const selectedModel = installed.has(current)
+        ? current
+        : installed.has(def)
+          ? def
+          : models[0]?.name ?? "";
+      set({ models, selectedModel });
     } finally {
       set({ loadingModels: false });
     }
@@ -180,13 +187,25 @@ export const useStore = create<LokiState>((set, get) => ({
 
   newSession: async () => {
     const s = await createSession(get().selectedModel || undefined);
-    set({ currentSessionId: s.id, messages: [], streamContent: "" });
+    set({
+      currentSessionId: s.id,
+      messages: [],
+      streamContent: "",
+      streamStatus: "",
+      streamNotice: null,
+    });
     await get().refreshSessions();
   },
 
   openSession: async (id) => {
     const { messages } = await getSession(id);
-    set({ currentSessionId: id, messages, streamContent: "" });
+    set({
+      currentSessionId: id,
+      messages,
+      streamContent: "",
+      streamStatus: "",
+      streamNotice: null,
+    });
   },
 
   removeSession: async (id) => {
@@ -201,13 +220,37 @@ export const useStore = create<LokiState>((set, get) => ({
     if (get().streaming) return;
     content = content.trim();
     if (!content) return;
+    if (!get().selectedModel) {
+      const errMsg: Message = {
+        id: `err-${Date.now()}`,
+        session_id: get().currentSessionId ?? "",
+        role: "assistant",
+        content: "⚠️ Aucun modèle Ollama installé ou sélectionné.",
+        created_at: Date.now() / 1000,
+      };
+      set({ messages: [...get().messages, errMsg] });
+      return;
+    }
 
     // Crée une session à la volée si aucune n'est ouverte.
     let sid = get().currentSessionId;
-    if (!sid) {
-      const s = await createSession(get().selectedModel || undefined);
-      sid = s.id;
-      set({ currentSessionId: s.id });
+    try {
+      if (!sid) {
+        const s = await createSession(get().selectedModel || undefined);
+        sid = s.id;
+        set({ currentSessionId: s.id });
+      }
+    } catch (err) {
+      const detail = err instanceof Error ? err.message : "backend injoignable";
+      const errMsg: Message = {
+        id: `err-${Date.now()}`,
+        session_id: "",
+        role: "assistant",
+        content: `⚠️ Impossible de créer la session : ${detail}`,
+        created_at: Date.now() / 1000,
+      };
+      set({ messages: [...get().messages, errMsg] });
+      return;
     }
 
     // Affichage optimiste du message utilisateur.
@@ -222,6 +265,8 @@ export const useStore = create<LokiState>((set, get) => ({
       messages: [...get().messages, userMsg],
       streaming: true,
       streamContent: "",
+      streamStatus: "Connexion à Ollama…",
+      streamNotice: null,
       streamTools: [],
       pendingShell: null,
     });
@@ -230,6 +275,8 @@ export const useStore = create<LokiState>((set, get) => ({
       { session_id: sid, content, model: get().selectedModel || undefined },
       {
         onToken: (t) => set({ streamContent: get().streamContent + t }),
+        onStatus: (message) => set({ streamStatus: message }),
+        onNotice: (message) => set({ streamNotice: message }),
         onToolCall: (call) =>
           set({ streamTools: [...get().streamTools, call] }),
         onToolResult: (call) => {
@@ -254,7 +301,13 @@ export const useStore = create<LokiState>((set, get) => ({
                 typeof t.args?.path === "string" &&
                 /\.html?$/.test(t.args.path as string)
             );
-          set({ streaming: false, streamContent: "", streamTools: [] });
+          set({
+            streaming: false,
+            streamContent: "",
+            streamStatus: "",
+            streamNotice: null,
+            streamTools: [],
+          });
           // Recharge depuis la base + l'arborescence (fichiers créés par l'agent).
           if (get().currentSessionId === sid) await get().openSession(sid!);
           await get().refreshSessions();
@@ -272,6 +325,8 @@ export const useStore = create<LokiState>((set, get) => ({
           set({
             streaming: false,
             streamContent: "",
+            streamStatus: "",
+            streamNotice: null,
             streamTools: [],
             messages: [...get().messages, errMsg],
           });