fix: assistant IA toujours en mode agent, retrait du bouton toggle #187

- Bouton « mode agent » du panneau supprimé : l'assistant est toujours
  agent (toute requête texte part sur /api/ai/bookslm/agent, les images
  restent sur /chat multimodal). Les mutations gardent la confirmation
  two-step. Deep Research et les quick actions `agent: true` basculaient
  déjà le mode en silence : le toggle ne protégeait plus rien.
- /agent résout le provider comme /chat (req.provider brut transmis à
  l'adapter pouvait désigner un fournisseur indisponible et retomber
  silencieusement sur un autre que l'étiquette SSE affichée).
- Schémas des outils mis en cache par (scope, taille du registre) avec
  copies fraîches par appelant (~27 ms de pydantic économisées par
  requête agent/MCP).
- Aide in-app réécrite, i18n FR/EN (retrait de ai.agent_mode_*),
  tests frontend adaptés (ai.test.mjs 100/100) + non-régression pytest
  (provider résolu, cache sûr).

backend/bookslm_routes.py

@@ -536,18 +536,6 @@ async def api_bookslm_agent(
         # run no longer pauses on every subsequent mutating call.
         ctx.confirmed = True
 
-    async def _llm(msgs, tool_schemas):
-        return await chat_completion(
-            msgs,
-            tools=tool_schemas,
-            provider=req.provider,
-            model=req.model,
-            temperature=0.3,
-            # Tool-call arguments can carry a whole file body (e.g. a generated
-            # table): leave more room than the plain-chat default.
-            max_tokens=8192,
-        )
-
     async def generate_sse():
         import asyncio
 
@@ -561,6 +549,23 @@ async def api_bookslm_agent(
                 yield f"event: error\ndata: {error_data}\n\n"
                 return
 
+            # #187: resolve the provider like /chat does — the agent must use
+            # the same engine the SSE "provider" tag reports (the raw
+            # req.provider could name an unavailable provider and silently
+            # fall back to another one via _get_provider_config).
+            async def _llm(msgs, tool_schemas):
+                return await chat_completion(
+                    msgs,
+                    tools=tool_schemas,
+                    provider=cfg_name,
+                    model=req.model,
+                    temperature=0.3,
+                    # Tool-call arguments can carry a whole file body (e.g. a
+                    # generated table): leave more room than the plain-chat
+                    # default.
+                    max_tokens=8192,
+                )
+
             # Stream tool events live: each executed step is pushed on the
             # queue by the loop callback and emitted as soon as it happens,
             # so the UI can grow its « N steps » block while thinking.

backend/tools/registry.py

@@ -117,9 +117,22 @@ def list_tools(*, scope: ToolScope | None = None) -> list[ToolSpec]:
     return specs
 
 
+# ponytail: tool schemas are static after import (registration is decorator
+# only); keying the cache on len(_REGISTRY) invalidates it if a tool is ever
+# registered at runtime. Rebuilding 50 pydantic JSON schemas cost ~27 ms per
+# agent/MCP request.
+_SCHEMAS_CACHE: dict[Any, list[dict[str, Any]]] = {}
+
+
 def get_tool_schemas(*, scope: ToolScope | None = None) -> list[dict[str, Any]]:
-    """Return OpenAI-compatible schemas for registered tools."""
-    return [spec.openai_schema() for spec in list_tools(scope=scope)]
+    """Return OpenAI-compatible schemas for registered tools (cached)."""
+    key = (scope, len(_REGISTRY))
+    cached = _SCHEMAS_CACHE.get(key)
+    if cached is None:
+        cached = [spec.openai_schema() for spec in list_tools(scope=scope)]
+        _SCHEMAS_CACHE.clear()
+        _SCHEMAS_CACHE[key] = cached
+    return [dict(s) for s in cached]
 
 
 def _audit(ctx: ToolContext, spec: ToolSpec, arguments: dict[str, Any], *, ok: bool, error: str | None = None) -> None: