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 OpenAI-compatible schemas for registered tools (cached)."""
-    return [spec.openai_schema() for spec in list_tools(scope=scope)]
+    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:

frontend/index.html

@@ -4674,9 +4674,9 @@
                                     conversation.
                                 </li>
                                 <li data-i18n="help.assistant_agent">
-                                    Le bouton « mode agent » active les outils (lire, lister,
+                                    L'assistant utilise toujours ses outils (lire, lister,
-                                    chercher) ; les actions de modification demandent une
+                                    chercher, modifier) ; les actions de modification demandent
-                                    confirmation avec aperçu des changements.
+                                    une confirmation avec aperçu des changements.
                                 </li>
                                 <li data-i18n="help.assistant_agent_run">
                                     Les actions s'affichent dans le fil : l'assistant regroupe les

frontend/js/ai-quick-actions.js

@@ -46,8 +46,8 @@ export const ACTION_CATALOG = Object.freeze([
   { id: 'checklist', cat: 'structure', icon: 'list-checks', labelKey: 'qa.checklist', promptKey: 'qa.checklist.prompt' },
   { id: 'plan', cat: 'structure', icon: 'list-ordered', labelKey: 'qa.plan', promptKey: 'qa.plan.prompt' },
   { id: 'memo', cat: 'structure', icon: 'scroll-text', labelKey: 'qa.memo', promptKey: 'qa.memo.prompt' },
-  { id: 'frontmatter', cat: 'structure', icon: 'braces', labelKey: 'qa.frontmatter', promptKey: 'qa.frontmatter.prompt', agent: true },
+  { id: 'frontmatter', cat: 'structure', icon: 'braces', labelKey: 'qa.frontmatter', promptKey: 'qa.frontmatter.prompt' },
-  { id: 'frontmatter_update', cat: 'structure', icon: 'refresh-cw', labelKey: 'qa.frontmatter_update', promptKey: 'qa.frontmatter_update.prompt', agent: true },
+  { id: 'frontmatter_update', cat: 'structure', icon: 'refresh-cw', labelKey: 'qa.frontmatter_update', promptKey: 'qa.frontmatter_update.prompt' },
   { id: 'backlinks', cat: 'structure', icon: 'link-2', labelKey: 'qa.backlinks', promptKey: 'qa.backlinks.prompt' },
   { id: 'sections', cat: 'structure', icon: 'heading', labelKey: 'qa.sections', promptKey: 'qa.sections.prompt' },
   // ── Code & Scripts ───────────────────────────────────────────────────

frontend/js/bookslm.js

@@ -151,9 +151,8 @@ class BooksLM {
     this._sessions = [];
     this._currentSessionId = null;
     this._pendingNewSession = false;
-    // Agent mode: routes chat through /api/ai/bookslm/agent so the model can
+    // #187: the assistant is always in agent mode (tools on /agent endpoint);
-    // call read/search tools and propose mutations (confirmation cards).
+    // the former header toggle was removed — images still use the /chat path.
-    this._agentMode = this._readAgentMode();
     // Ad-hoc context added with `@` (files/directories) and images attached
     // by paste or by mentioning an image file.
     this._adhocFiles = [];
@@ -188,33 +187,6 @@ class BooksLM {
     this._extensions = null;
   }
 
-  _readAgentMode() {
-    try {
-      return localStorage.getItem('obsigate-bookslm-agent') === 'true';
-    } catch {
-      return false;
-    }
-  }
-
-  _toggleAgentMode() {
-    this._agentMode = !this._agentMode;
-    try {
-      localStorage.setItem('obsigate-bookslm-agent', this._agentMode ? 'true' : 'false');
-    } catch { /* private mode */ }
-    this._updateAgentToggle();
-  }
-
-  _updateAgentToggle() {
-    if (!this._panel) return;
-    const btn = this._panel.querySelector('.bookslm-btn-agent');
-    if (!btn) return;
-    btn.classList.toggle('active', this._agentMode);
-    const label = this._agentMode ? t('ai.agent_mode_on') : t('ai.agent_mode_off');
-    btn.title = label;
-    btn.setAttribute('aria-label', label);
-    btn.setAttribute('aria-pressed', this._agentMode ? 'true' : 'false');
-  }
-
   // ── Public API ──────────────────────────────────────────────────────
 
   /**
@@ -966,7 +938,6 @@ class BooksLM {
         </div>
         <span class="bookslm-qa-badge hidden" data-qa-context=""></span>
         <div class="bookslm-header-actions">
-          <button class="bookslm-btn-agent" title="${t('ai.agent_mode_off')}" aria-label="${t('ai.agent_mode_off')}" aria-pressed="false"><i data-lucide="bot" style="width:16px;height:16px"></i></button>
           <button class="bookslm-btn-history" title="${t('bookslm.session_history')}" aria-label="${t('bookslm.session_history')}"><i data-lucide="history" style="width:16px;height:16px"></i></button>
           <button class="bookslm-btn-new" title="${t('bookslm.new_conversation')}" aria-label="${t('bookslm.new_conversation')}"><i data-lucide="plus" style="width:16px;height:16px"></i></button>
           <button class="bookslm-btn-export" title="${t('bookslm.export')}" aria-label="${t('bookslm.export')}"><i data-lucide="download" style="width:16px;height:16px"></i></button>
@@ -1011,7 +982,6 @@ class BooksLM {
 
     panel.querySelector('.bookslm-btn-close').addEventListener('click', () => this.close());
     panel.querySelector('.bookslm-btn-new').addEventListener('click', () => this.newConversation());
-    panel.querySelector('.bookslm-btn-agent').addEventListener('click', () => this._toggleAgentMode());
     panel.querySelector('.bookslm-btn-history').addEventListener('click', (e) => {
       e.stopPropagation();
       this._toggleHistoryMenu();
@@ -1552,10 +1522,9 @@ class BooksLM {
     if (typeof safeCreateIcons === 'function') safeCreateIcons();
   }
 
-  /** #97/#99 — Deep Research: agent tools + a skill-like chip (no composer text). */
+  /** #97/#99 — Deep Research: a skill-like chip (the agent tools are always on). */
   _startDeepResearch() {
     this._closeExtMenu();
-    if (!this._agentMode) this._toggleAgentMode();
     this._activeDeepResearch = true;
     this._renderAttachments();
     showToast(t('bookslm.deep_research_started'), 'info');
@@ -2201,15 +2170,14 @@ class BooksLM {
   }
 
   /** Immediate-send the action prompt through the composer (same path as a
-   *  typed message: skills, agent mode and images all keep working). Actions
+   *  typed message: skills and images all keep working). #187: the assistant
-   *  flagged `agent` (they mutate the document) transparently switch the
+   *  is always agent mode, so the former `agent: true` quick actions need no
-   *  assistant to agent mode first — same pattern as Deep Research. */
+   *  mode switch anymore. */
   _runQuickAction(action) {
     if (!this._panel || !action) return;
     const { prompt } = actionTexts(action);
     if (!prompt) return;
     this._closeActionDrawer();
-    if (action.agent && !this._agentMode) this._toggleAgentMode();
     const textarea = this._panel.querySelector('textarea');
     if (!textarea) return;
     textarea.value = prompt;
@@ -3253,14 +3221,15 @@ class BooksLM {
     this._saveHistory();
   }
 
-  /** POST to /chat or /agent depending on the agent-mode toggle. */
+  /** POST to /agent (always) or /chat when images are attached. */
   _postChat(payload) {
-    // Images require the plain multimodal chat endpoint (the agent loop is
+    // #187: the assistant is always agent mode — images are the only case
+    // that still needs the plain multimodal chat endpoint (the agent loop is
     // text/tool oriented).
     const hasImages = Array.isArray(payload.images) && payload.images.length > 0;
-    const endpoint = (this._agentMode && !hasImages)
+    const endpoint = hasImages
-      ? '/api/ai/bookslm/agent'
+      ? '/api/ai/bookslm/chat'
-      : '/api/ai/bookslm/chat';
+      : '/api/ai/bookslm/agent';
     const headers = { 'Content-Type': 'application/json', ...(AuthManager.getAuthHeaders() || {}) };
     return fetch(endpoint, {
       method: 'POST',

tests/frontend/ai.test.mjs

@@ -159,7 +159,8 @@ async function main() {
       read: async () => (chunks.length ? { done: false, value: chunks.shift() } : { done: true }),
     };
     globalThis.fetch = async (url, opts) => {
-      if (String(url).includes("/api/ai/bookslm/chat")) {
+      // #187: plain-text sends now go to /agent — accept both bookslm endpoints.
+      if (/\/api\/ai\/bookslm\/(chat|agent)/.test(String(url))) {
         capturedBody = JSON.parse(opts.body);
         return { ok: true, status: 200, body: { getReader: () => reader } };
       }
@@ -1171,38 +1172,28 @@ async function main() {
     assert.equal(b._messages[0].content, "Ancienne conversation");
   });
 
-  // ── 14. Agent mode & confirmations (B5) ──
+  // ── 14. Agent mode & confirmations (B5 / #187) ──
-  await test("agent mode toggle persists and switches the endpoint", async () => {
+  await test("#187: no agent toggle — plain requests always target /agent", async () => {
     localStorage.clear();
     const b = new BooksLM();
-    assert.equal(b._agentMode, false, "agent mode off by default");
+    assert.equal(b._agentMode, undefined, "agent-mode state removed");
-    b._toggleAgentMode();
+    assert.equal(b._toggleAgentMode, undefined, "toggle method removed");
-    assert.equal(b._agentMode, true);
-    assert.equal(localStorage.getItem("obsigate-bookslm-agent"), "true");
-    const b2 = new BooksLM();
-    assert.equal(b2._agentMode, true, "persisted across instances");
 
     let url = null;
     globalThis.fetch = async (u) => { url = String(u); return { ok: true, status: 200, json: async () => ({}) }; };
-    b2._abortCtrl = null;
+    b._abortCtrl = null;
-    await b2._postChat({ message: "x" });
+    await b._postChat({ message: "x" });
-    assert.ok(url.includes("/api/ai/bookslm/agent"), "agent mode targets /agent");
+    assert.ok(url.includes("/api/ai/bookslm/agent"), "default targets /agent");
-    b2._agentMode = false;
+    assert.equal(localStorage.getItem("obsigate-bookslm-agent"), null, "no legacy persistence");
-    await b2._postChat({ message: "x" });
-    assert.ok(url.includes("/api/ai/bookslm/chat"), "default targets /chat");
     localStorage.clear();
   });
 
-  await test("header exposes the agent-mode toggle", () => {
+  await test("header no longer exposes the agent-mode toggle", () => {
     const b = new BooksLM();
     const panel = b._render();
     b._panel = panel;
     document.body.appendChild(panel);
-    const btn = panel.querySelector(".bookslm-btn-agent");
+    assert.equal(panel.querySelector(".bookslm-btn-agent"), null, "toggle button removed");
-    assert.ok(btn, "agent toggle button present");
-    assert.equal(btn.getAttribute("aria-pressed"), "false");
-    btn.click();
-    assert.equal(btn.getAttribute("aria-pressed"), "true");
     panel.remove();
     localStorage.clear();
   });
@@ -1236,7 +1227,6 @@ async function main() {
     };
 
     const b = new BooksLM();
-    b._agentMode = true;
     b._panel = b._render();
     document.body.appendChild(b._panel);
     const msg = {
@@ -1418,7 +1408,7 @@ async function main() {
 
     let posted = null;
     globalThis.fetch = async (url, opts) => {
-      if (String(url).includes("/api/ai/bookslm/chat")) posted = JSON.parse(opts.body);
+      if (/\/api\/ai\/bookslm\/(chat|agent)/.test(String(url))) posted = JSON.parse(opts.body);
       return { ok: true, status: 200, body: { getReader: () => ({ read: async () => ({ done: true }) }) } };
     };
     await b._sendMessage();
@@ -1432,10 +1422,9 @@ async function main() {
     localStorage.clear();
   });
 
-  // ── 32. Images force the plain chat endpoint (not the agent) (#81) ──
+  // ── 32. Images force the plain chat endpoint (#81 / #187: always-on agent) ──
-  await test("images route to /chat even in agent mode", async () => {
+  await test("images route to /chat, text to /agent", async () => {
     const b = new BooksLM();
-    b._agentMode = true;
     let url = "";
     globalThis.fetch = async (u) => {
       url = String(u);

tests/test_bookslm.py

@@ -940,6 +940,31 @@ class TestBooksLMAgentEndpoint:
         assert "create_file" in captured["system"]
         assert '"action": "create_file"' not in captured["system"]
 
+    def test_agent_llm_uses_resolved_provider(self, bookslm_client, monkeypatch):
+        """#187: the agent loop must call the provider resolved by
+        _resolve_provider_name (not the raw request value), so the provider
+        tag reported by SSE matches the engine that actually answered."""
+        import backend.bookslm_routes as routes
+        from backend.ai_chat import LLMResponse
+
+        seen = {}
+
+        async def fake_chat_completion(messages, **kwargs):
+            seen["provider"] = kwargs.get("provider")
+            return LLMResponse(content="ok")
+
+        monkeypatch.setattr(routes, "chat_completion", fake_chat_completion)
+        monkeypatch.setattr(routes, "_resolve_provider_name", lambda requested: "openrouter")
+
+        token, _ = _login_bookslm(bookslm_client)
+        resp = bookslm_client.post(
+            "/api/ai/bookslm/agent",
+            json={"directory": "", "message": "salut", "mode": "general", "provider": "nope"},
+            headers={"Authorization": f"Bearer {token}"},
+        )
+        assert resp.status_code == 200
+        assert seen["provider"] == "openrouter"
+
     def test_agent_tool_call_flow(self, bookslm_client, monkeypatch):
         import backend.bookslm_routes as routes
         from backend.ai_chat import LLMResponse, ToolCall

tests/test_tools.py

@@ -93,6 +93,16 @@ class TestRegistry:
         props = read["function"]["parameters"]["properties"]
         assert "vault" in props and "path" in props
 
+    def test_schemas_cached_but_caller_safe(self):
+        """#187: repeated calls hit the cache and a caller mutating its copy
+        (e.g. stripping keys) must never corrupt the next request."""
+        a = get_tool_schemas()
+        b = get_tool_schemas()
+        assert a == b
+        assert all(x is not y for x, y in zip(a, b)), "top-level dicts must be fresh copies"
+        a[0]["poisoned"] = True
+        assert "poisoned" not in get_tool_schemas()[0]
+
     def test_duplicate_tool_name_raises(self):
         from backend.tools.registry import tool