fix(ai): garantir une reponse finale quand la boucle d'agent epuise son budget (BUG-052)

backend/agent/loop.py

@@ -40,6 +40,15 @@ MAX_TOOL_RESULT_CHARS = 100_000
 # Quota: maximum tool calls executed per agent run (``BOOKSLM_MAX_TOOL_CALLS``).
 DEFAULT_MAX_TOOL_CALLS = int(os.environ.get("BOOKSLM_MAX_TOOL_CALLS", "25"))
 
+# Sent as a last user turn when the loop stopped before the model produced an
+# answer (iteration/quota budget exhausted while it was still calling tools).
+_FINALIZE_INSTRUCTION = (
+    "N'appelle plus aucun outil. Réponds maintenant directement à l'utilisateur, "
+    "en français, à partir des informations déjà recueillies ci-dessus. "
+    "Structure la réponse en Markdown, cite les liens sources utiles, et si les "
+    "informations sont insuffisantes, dis-le explicitement."
+)
+
 # Stopping reasons
 STOP_DONE = "done"
 STOP_MAX_ITERATIONS = "max_iterations"
@@ -109,8 +118,8 @@ def _assistant_tool_message(content: str | None, tool_calls: list[Any]) -> dict[
     }
 
 
-def _deferred_tool_message(call: Any) -> dict[str, Any]:
-    """Answer a tool call that was not reached because the run paused.
+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
@@ -125,7 +134,7 @@ def _deferred_tool_message(call: Any) -> dict[str, Any]:
         "name": call.name,
         "content": json.dumps({
             "status": "deferred",
-            "reason": (
+            "reason": reason or (
                 "Not executed: the run paused to confirm an earlier tool call. "
                 "Re-issue this call if it is still needed."
             ),
@@ -133,6 +142,69 @@ def _deferred_tool_message(call: Any) -> dict[str, Any]:
     }
 
 
+def _fallback_summary(executed: list[ToolCallRecord]) -> str:
+    """Deterministic non-empty answer built from the gathered tool results.
+
+    Used only if the final synthesis call fails or returns nothing, so a turn
+    never ends on an empty message (BUG-052).
+    """
+    lines: list[str] = []
+    for record in executed:
+        data = record.result
+        if not isinstance(data, dict):
+            continue
+        for item in (data.get("results") or [])[:5]:
+            if not isinstance(item, dict):
+                continue
+            title = item.get("title") or item.get("url") or ""
+            url = item.get("url") or ""
+            lines.append(f"- [{title}]({url})" if url else f"- {title}")
+        if data.get("url") and data.get("text"):
+            title = data.get("title") or data["url"]
+            lines.append(f"- [{title}]({data['url']})")
+    if not lines:
+        return "Je n'ai pas pu produire de réponse à partir des résultats obtenus."
+    unique = list(dict.fromkeys(lines))
+    return "Voici les sources pertinentes trouvées :\n" + "\n".join(unique)
+
+
+async def _finalize_answer(
+    llm: Callable[..., Any],
+    convo: list[dict[str, Any]],
+    executed: list[ToolCallRecord],
+    steps: list[dict[str, Any]],
+    iterations: int,
+    stopped: str,
+) -> AgentResult:
+    """Guarantee a textual answer when the loop stopped before producing one.
+
+    Web research often exhausts the iteration budget while the model is still
+    calling tools; returning ``content=""`` left the conversation with steps and
+    sources but no answer. One final tool-less call asks the model to synthesize
+    the gathered results, and a deterministic source list is used as a last
+    resort (BUG-052).
+    """
+    content = ""
+    if executed:
+        try:
+            response = await llm(
+                [*convo, {"role": "user", "content": _FINALIZE_INSTRUCTION}], []
+            )
+            content = (response.content or "").strip()
+        except Exception as e:
+            logger.warning(f"Agent final synthesis failed: {e}")
+        if not content:
+            content = _fallback_summary(executed)
+    return AgentResult(
+        content=content,
+        messages=convo,
+        tool_calls=executed,
+        steps=steps,
+        iterations=iterations,
+        stopped=stopped,
+    )
+
+
 def _execute_confirmed(
     ctx: ToolContext,
     confirm_pending: dict[str, Any],
@@ -275,13 +347,14 @@ async def run_agent(
         for index, call in enumerate(response.tool_calls):
             if quota is not None and len(executed) >= quota:
                 logger.warning(f"Agent reached the tool-call quota ({quota})")
-                return AgentResult(
-                    content=response.content or "",
-                    messages=convo,
-                    tool_calls=executed,
-                    steps=steps,
-                    iterations=iteration,
-                    stopped=STOP_QUOTA_EXCEEDED,
+                # Keep the conversation valid for the synthesis call: the
+                # assistant message announced every tool call of the batch.
+                for skipped in response.tool_calls[index:]:
+                    convo.append(_deferred_tool_message(
+                        skipped, "Not executed: the tool-call quota was reached."
+                    ))
+                return await _finalize_answer(
+                    llm, convo, executed, steps, iteration, STOP_QUOTA_EXCEEDED
                 )
             try:
                 result = call_tool(call.name, ctx, call.arguments)
@@ -327,11 +400,6 @@ async def run_agent(
             })
 
     logger.warning(f"Agent reached max iterations ({max_iterations})")
-    return AgentResult(
-        content="",
-        messages=convo,
-        tool_calls=executed,
-        steps=steps,
-        iterations=max_iterations,
-        stopped=STOP_MAX_ITERATIONS,
+    return await _finalize_answer(
+        llm, convo, executed, steps, max_iterations, STOP_MAX_ITERATIONS
     )