feat: assistant IA — approbation groupee des actions, bloc d'etapes, refresh UI et bouton Stop (BUG-074, BUG-075, BUG-076, BUG-077)

backend/agent/loop.py

@@ -28,6 +28,7 @@ from backend.tools.api import (
     ToolError,
     ToolScope,
     call_tool,
+    get_tool,
     get_tool_schemas,
 )
 from backend.tools.labels import thought_step_label, tool_step_label
@@ -121,12 +122,15 @@ def _assistant_tool_message(content: str | None, tool_calls: list[Any]) -> dict[
 def _deferred_tool_message(call: Any, reason: str | None = None) -> dict[str, Any]:
     """Answer a tool call that was not reached because the run stopped early.
 
-    A single LLM response may carry several tool calls. When one of them is
-    mutating and pauses the run for confirmation, the assistant message already
-    lists *all* of them, so every ``tool_call_id`` must get a tool result before
-    the next LLM call (the OpenAI tool protocol rejects dangling ids). The calls
-    that were not reached get a synthetic ``deferred`` result; the model
-    re-issues them once the confirmed call has been applied (BUG-050).
+    A single LLM response may carry several tool calls; when the run stops
+    before reaching some of them (tool-call quota), the assistant message still
+    lists *all* of them, so every ``tool_call_id`` must get a tool result
+    before the next LLM call (the OpenAI tool protocol rejects dangling ids).
+    The calls that were not reached get a synthetic ``deferred`` result.
+
+    Note: mutating calls that pause the run for confirmation are no longer
+    deferred — they are batched and applied together on resume (BUG-075); this
+    helper remains for budget stops (BUG-050/BUG-052).
     """
     return {
         "role": "tool",
@@ -135,13 +139,29 @@ def _deferred_tool_message(call: Any, reason: str | None = None) -> dict[str, An
         "content": json.dumps({
             "status": "deferred",
             "reason": reason or (
-                "Not executed: the run paused to confirm an earlier tool call. "
+                "Not executed: the run stopped before reaching this tool call. "
                 "Re-issue this call if it is still needed."
             ),
         }, ensure_ascii=False),
     }
 
 
+def _action_descriptor(call: Any) -> dict[str, Any]:
+    """Describe one paused mutating tool call for the confirmation payload.
+
+    A single LLM response may request several mutations (create a folder and
+    the files inside it…). They are batched into one confirmation so the user
+    approves the whole plan in one click (BUG-075). ``step`` reuses the
+    Notion-style label, so the confirmation card reads like the steps block.
+    """
+    return {
+        "id": call.id,
+        "tool": call.name,
+        "arguments": call.arguments,
+        "step": tool_step_label(call.name, call.arguments),
+    }
+
+
 def _fallback_summary(executed: list[ToolCallRecord]) -> str:
     """Deterministic non-empty answer built from the gathered tool results.
 
@@ -212,53 +232,66 @@ def _execute_confirmed(
     executed: list[ToolCallRecord],
     on_tool_call: Callable[[ToolCallRecord], None] | None,
 ) -> None:
-    """Apply a previously-paused mutating tool call and feed its result back.
+    """Apply previously-paused mutating tool calls and feed their results back.
 
     The pending payload is the ``error`` object emitted by a ``confirmation``
-    event. The assistant tool-call message is expected to already be in
+    event, optionally carrying an ``actions`` list with every mutating call of
+    the LLM turn (BUG-075). Each action is applied with a one-shot confirmation
+    and its ``tool_call_id`` answered, keeping the conversation valid for the
+    resumed turn. The assistant tool-call message is expected to already be in
     ``convo`` (it is part of the snapshot returned with the confirmation).
     """
     from backend.ai_chat import ToolCall
 
-    error = confirm_pending.get("error", confirm_pending)
-    name = error.get("tool")
-    arguments = error.get("arguments") or {}
-    call_id = error.get("id") or "call_pending"
+    error = confirm_pending.get("error", confirm_pending) or {}
+    actions = confirm_pending.get("actions")
+    if not isinstance(actions, list) or not actions:
+        # Legacy single-action payload (no ``actions`` list).
+        actions = [{
+            "id": error.get("id") or "call_pending",
+            "tool": error.get("tool"),
+            "arguments": error.get("arguments") or {},
+        }]
 
-    if not name:
-        raise ToolError("Malformed confirmation payload", code="invalid_confirmation")
+    for action in actions:
+        name = action.get("tool")
+        arguments = action.get("arguments") or {}
+        call_id = action.get("id") or "call_pending"
 
-    # Make sure the assistant tool-call message is present in the snapshot.
-    if not any(
-        m.get("role") == "assistant" and any(
-            tc.get("id") == call_id for tc in (m.get("tool_calls") or [])
+        if not name:
+            raise ToolError("Malformed confirmation payload", code="invalid_confirmation")
+
+        # Make sure the assistant tool-call message is present in the snapshot.
+        if not any(
+            m.get("role") == "assistant" and any(
+                tc.get("id") == call_id for tc in (m.get("tool_calls") or [])
+            )
+            for m in convo
+        ):
+            convo.append(_assistant_tool_message(None, [ToolCall(id=call_id, name=name, arguments=arguments)]))
+
+        try:
+            result = call_tool(name, ctx, arguments, confirm=True)
+            payload = result.data
+            ok = True
+        except ToolError as e:
+            payload = e.to_dict()
+            ok = False
+
+        record = ToolCallRecord(
+            name=name, arguments=arguments, ok=ok, result=payload,
+            step=tool_step_label(name, arguments),
         )
-        for m in convo
-    ):
-        convo.append(_assistant_tool_message(None, [ToolCall(id=call_id, name=name, arguments=arguments)]))
+        executed.append(record)
+        if on_tool_call is not None:
+            on_tool_call(record)
 
-    try:
-        result = call_tool(name, ctx, arguments, confirm=True)
-        payload = result.data
-        ok = True
-    except ToolError as e:
-        payload = e.to_dict()
-        ok = False
-
-    record = ToolCallRecord(
-        name=name, arguments=arguments, ok=ok, result=payload,
-        step=tool_step_label(name, arguments),
-    )
-    executed.append(record)
-    if on_tool_call is not None:
-        on_tool_call(record)
-
-    convo.append({
-        "role": "tool",
-        "tool_call_id": call_id,
-        "name": name,
-        "content": json.dumps(_truncate(payload), ensure_ascii=False, default=str),
-    })
+        convo.append({
+            "role": "tool",
+            "tool_call_id": call_id,
+            "name": name,
+            "content": json.dumps(_truncate(payload), ensure_ascii=False, default=str),
+        })
 
 
 async def run_agent(
@@ -315,6 +348,36 @@ async def run_agent(
     convo = [dict(m) for m in (resume_messages if resume_messages is not None else messages)]
     executed: list[ToolCallRecord] = []
 
+    def _run_call(call: Any) -> None:
+        """Execute one tool call, record it and answer its ``tool_call_id``.
+
+        ``ToolConfirmationRequired`` propagates to the caller so the loop can
+        pause and batch the mutating calls of the turn (BUG-075).
+        """
+        try:
+            result = call_tool(call.name, ctx, call.arguments)
+            payload: Any = result.data
+            ok = True
+        except ToolConfirmationRequired:
+            raise
+        except ToolError as e:
+            payload = e.to_dict()
+            ok = False
+        record = ToolCallRecord(
+            name=call.name, arguments=call.arguments, ok=ok, result=payload,
+            step=tool_step_label(call.name, call.arguments),
+        )
+        executed.append(record)
+        steps.append(record.step)
+        if on_tool_call is not None:
+            on_tool_call(record)
+        convo.append({
+            "role": "tool",
+            "tool_call_id": call.id,
+            "name": call.name,
+            "content": json.dumps(_truncate(payload), ensure_ascii=False, default=str),
+        })
+
     if confirm_pending:
         if quota is not None and len(executed) >= quota:
             return AgentResult(
@@ -357,19 +420,25 @@ async def run_agent(
                     llm, convo, executed, steps, iteration, STOP_QUOTA_EXCEEDED
                 )
             try:
-                result = call_tool(call.name, ctx, call.arguments)
-                payload = result.data
-                ok = True
+                _run_call(call)
             except ToolConfirmationRequired as e:
                 logger.info(f"Agent paused: confirmation required for '{call.name}'")
                 pending = e.to_dict()
                 # Include the tool-call id so the client can echo it back.
                 pending["error"]["id"] = call.id
-                # BUG-050: the assistant message lists every tool call of this
-                # batch, so answer the ones we did not reach to keep the
-                # conversation valid for the resumed turn.
-                for skipped in response.tool_calls[index + 1:]:
-                    convo.append(_deferred_tool_message(skipped))
+                # BUG-075: batch every mutating call of this LLM turn so the
+                # user approves the whole plan at once (one resume applies them
+                # all) instead of approving one action after another. Read-only
+                # calls of the batch run immediately and answer their
+                # ``tool_call_id`` so the resumed turn stays valid.
+                actions = [_action_descriptor(call)]
+                for after in response.tool_calls[index + 1:]:
+                    spec = get_tool(after.name)
+                    if spec is not None and spec.requires_confirmation:
+                        actions.append(_action_descriptor(after))
+                    else:
+                        _run_call(after)
+                pending["actions"] = actions
                 return AgentResult(
                     content=response.content or "",
                     messages=convo,
@@ -379,25 +448,6 @@ async def run_agent(
                     stopped=STOP_CONFIRMATION_REQUIRED,
                     pending=pending,
                 )
-            except ToolError as e:
-                payload = e.to_dict()
-                ok = False
-
-            record = ToolCallRecord(
-                name=call.name, arguments=call.arguments, ok=ok, result=payload,
-                step=tool_step_label(call.name, call.arguments),
-            )
-            executed.append(record)
-            steps.append(record.step)
-            if on_tool_call is not None:
-                on_tool_call(record)
-
-            convo.append({
-                "role": "tool",
-                "tool_call_id": call.id,
-                "name": call.name,
-                "content": json.dumps(_truncate(payload), ensure_ascii=False, default=str),
-            })
 
     logger.warning(f"Agent reached max iterations ({max_iterations})")
     return await _finalize_answer(

backend/bookslm_routes.py

@@ -110,6 +110,12 @@ class BooksLMChatRequest(BaseModel):
         description="Conversation snapshot returned alongside a ``confirmation`` event, "
                     "echoed back to resume the agent run.",
     )
+    confirm_all: bool = Field(
+        default=False,
+        description="Global approval (BUG-075): apply every pending action of the batch "
+                    "and auto-approve the remaining mutating calls of the same run, "
+                    "so the run does not pause on each action.",
+    )
     app_context: dict[str, Any] | None = Field(
         default=None,
         description="Live client UI state for the General assistant: open_documents, "
@@ -506,9 +512,11 @@ async def api_bookslm_agent(
     Same context as ``/chat`` but the model may call tools (read/search the
     vault) through the shared tool layer. Emits one ``tool`` event per executed
     tool call, then a final ``message`` event. Mutating tools pause the run with
-    a ``confirmation`` event (two-step propose/apply) carrying the pending call
-    and the conversation snapshot; the client resumes by echoing them back in
-    ``confirm`` / ``confirm_messages``.
+    a ``confirmation`` event (two-step propose/apply) carrying the pending
+    ``actions`` (every mutating call of the turn) and the conversation snapshot;
+    the client resumes by echoing them back in ``confirm`` / ``confirm_messages``,
+    optionally with ``confirm_all`` to apply the whole batch and auto-approve the
+    rest of the run (BUG-075).
     """
     _validate_vision_support(req)
     system_prompt = _resolve_system_prompt(req, current_user, agent=True)
@@ -523,6 +531,10 @@ async def api_bookslm_agent(
     messages.append({"role": "user", "content": _build_user_content(req, vault_path)})
 
     ctx = ToolContext(user=current_user, mode=ToolMode.IN_APP)
+    if req.confirm_all:
+        # BUG-075: a single global approval authorizes the whole plan, so the
+        # run no longer pauses on every subsequent mutating call.
+        ctx.confirmed = True
 
     async def _llm(msgs, tool_schemas):
         return await chat_completion(