Better handling of LLM request failures

strix/agents/base_agent.py

@@ -13,7 +13,7 @@ from jinja2 import (
     select_autoescape,
 )
 
-from strix.llm import LLM, LLMConfig
+from strix.llm import LLM, LLMConfig, LLMRequestFailedError
 from strix.llm.utils import clean_content
 from strix.tools import process_tool_invocations
 
@@ -164,6 +164,10 @@ class BaseAgent(metaclass=AgentMeta):
                 await self._enter_waiting_state(tracer)
                 continue
 
+            if self.state.llm_failed:
+                await self._wait_for_input()
+                continue
+
             self.state.increment_iteration()
 
             try:
@@ -176,6 +180,13 @@ class BaseAgent(metaclass=AgentMeta):
                 await self._enter_waiting_state(tracer, error_occurred=False, was_cancelled=True)
                 continue
 
+            except LLMRequestFailedError as e:
+                self.state.add_error(f"LLM request failed: {e}")
+                self.state.enter_waiting_state(llm_failed=True)
+                if tracer:
+                    tracer.update_agent_status(self.state.agent_id, "llm_failed")
+                continue
+
             except (RuntimeError, ValueError, TypeError) as e:
                 if not await self._handle_iteration_error(e, tracer):
                     await self._enter_waiting_state(tracer, error_occurred=True)
@@ -327,7 +338,7 @@ class BaseAgent(metaclass=AgentMeta):
             tracer.update_agent_status(self.state.agent_id, "error")
         return True
 
-    def _check_agent_messages(self, state: AgentState) -> None:
+    def _check_agent_messages(self, state: AgentState) -> None:  # noqa: PLR0912
         try:
             from strix.tools.agents_graph.agents_graph_actions import _agent_graph, _agent_messages
 
@@ -340,13 +351,29 @@ class BaseAgent(metaclass=AgentMeta):
                 has_new_messages = False
                 for message in messages:
                     if not message.get("read", False):
-                        if state.is_waiting_for_input():
-                            state.resume_from_waiting()
-                            has_new_messages = True
-
-                        sender_name = "Unknown Agent"
                         sender_id = message.get("from")
 
+                        if state.is_waiting_for_input():
+                            if state.llm_failed:
+                                if sender_id == "user":
+                                    state.resume_from_waiting()
+                                    has_new_messages = True
+
+                                    from strix.cli.tracer import get_global_tracer
+
+                                    tracer = get_global_tracer()
+                                    if tracer:
+                                        tracer.update_agent_status(state.agent_id, "running")
+                            else:
+                                state.resume_from_waiting()
+                                has_new_messages = True
+
+                                from strix.cli.tracer import get_global_tracer
+
+                                tracer = get_global_tracer()
+                                if tracer:
+                                    tracer.update_agent_status(state.agent_id, "running")
+
                         if sender_id == "user":
                             sender_name = "User"
                             state.add_message("user", message.get("content", ""))

strix/agents/state.py

@@ -23,6 +23,7 @@ class AgentState(BaseModel):
     completed: bool = False
     stop_requested: bool = False
     waiting_for_input: bool = False
+    llm_failed: bool = False
     final_result: dict[str, Any] | None = None
 
     messages: list[dict[str, Any]] = Field(default_factory=list)
@@ -85,15 +86,17 @@ class AgentState(BaseModel):
     def is_waiting_for_input(self) -> bool:
         return self.waiting_for_input
 
-    def enter_waiting_state(self) -> None:
+    def enter_waiting_state(self, llm_failed: bool = False) -> None:
         self.waiting_for_input = True
         self.stop_requested = False
+        self.llm_failed = llm_failed
         self.last_updated = datetime.now(UTC).isoformat()
 
     def resume_from_waiting(self, new_task: str | None = None) -> None:
         self.waiting_for_input = False
         self.stop_requested = False
         self.completed = False
+        self.llm_failed = False
         if new_task:
             self.task = new_task
         self.last_updated = datetime.now(UTC).isoformat()