feat: add reset functionality to semantic belief extractor

ref: N25B-432

src/control_backend/agents/bdi/bdi_core_agent.py

@@ -205,12 +205,15 @@ class BDICoreAgent(BaseAgent):
 
         self.logger.debug(f"Added belief {self.format_belief_string(name, args)}")
 
-    def _remove_belief(self, name: str, args: Iterable[str]):
+    def _remove_belief(self, name: str, args: Iterable[str] | None):
         """
         Removes a specific belief (with arguments), if it exists.
         """
-        new_args = (agentspeak.Literal(arg) for arg in args)
-        term = agentspeak.Literal(name, new_args)
+        if args is None:
+            term = agentspeak.Literal(name)
+        else:
+            new_args = (agentspeak.Literal(arg) for arg in args)
+            term = agentspeak.Literal(name, new_args)
 
         result = self.bdi_agent.call(
             agentspeak.Trigger.removal,
@@ -346,8 +349,8 @@ class BDICoreAgent(BaseAgent):
         self.logger.info("Message sent to LLM agent: %s", text)
 
     @staticmethod
-    def format_belief_string(name: str, args: Iterable[str] = []):
+    def format_belief_string(name: str, args: Iterable[str] | None = []):
         """
         Given a belief's name and its args, return a string of the form "name(*args)"
         """
-        return f"{name}{'(' if args else ''}{','.join(args)}{')' if args else ''}"
+        return f"{name}{'(' if args else ''}{','.join(args or [])}{')' if args else ''}"

src/control_backend/agents/bdi/bdi_program_manager.py

@@ -7,9 +7,9 @@ from zmq.asyncio import Context
 from control_backend.agents import BaseAgent
 from control_backend.agents.bdi.agentspeak_generator import AgentSpeakGenerator
 from control_backend.core.config import settings
-from control_backend.schemas.belief_list import BeliefList
+from control_backend.schemas.belief_list import BeliefList, GoalList
 from control_backend.schemas.internal_message import InternalMessage
-from control_backend.schemas.program import Belief, ConditionalNorm, InferredBelief, Program
+from control_backend.schemas.program import Belief, ConditionalNorm, Goal, InferredBelief, Program
 
 
 class BDIProgramManager(BaseAgent):
@@ -63,24 +63,23 @@ class BDIProgramManager(BaseAgent):
     def _extract_beliefs_from_program(program: Program) -> list[Belief]:
         beliefs: list[Belief] = []
 
+        def extract_beliefs_from_belief(belief: Belief) -> list[Belief]:
+            if isinstance(belief, InferredBelief):
+                return extract_beliefs_from_belief(belief.left) + extract_beliefs_from_belief(
+                    belief.right
+                )
+            return [belief]
+
         for phase in program.phases:
             for norm in phase.norms:
                 if isinstance(norm, ConditionalNorm):
-                    beliefs += BDIProgramManager._extract_beliefs_from_belief(norm.condition)
+                    beliefs += extract_beliefs_from_belief(norm.condition)
 
             for trigger in phase.triggers:
-                beliefs += BDIProgramManager._extract_beliefs_from_belief(trigger.condition)
+                beliefs += extract_beliefs_from_belief(trigger.condition)
 
         return beliefs
 
-    @staticmethod
-    def _extract_beliefs_from_belief(belief: Belief) -> list[Belief]:
-        if isinstance(belief, InferredBelief):
-            return BDIProgramManager._extract_beliefs_from_belief(
-                belief.left
-            ) + BDIProgramManager._extract_beliefs_from_belief(belief.right)
-        return [belief]
-
     async def _send_beliefs_to_semantic_belief_extractor(self, program: Program):
         """
         Extract beliefs from the program and send them to the Semantic Belief Extractor Agent.
@@ -98,6 +97,46 @@ class BDIProgramManager(BaseAgent):
 
         await self.send(message)
 
+    @staticmethod
+    def _extract_goals_from_program(program: Program) -> list[Goal]:
+        """
+        Extract all goals from the program, including subgoals.
+
+        :param program: The program received from the API.
+        :return: A list of Goal objects.
+        """
+        goals: list[Goal] = []
+
+        def extract_goals_from_goal(goal_: Goal) -> list[Goal]:
+            goals_: list[Goal] = [goal]
+            for plan in goal_.plan:
+                if isinstance(plan, Goal):
+                    goals_.extend(extract_goals_from_goal(plan))
+            return goals_
+
+        for phase in program.phases:
+            for goal in phase.goals:
+                goals.extend(extract_goals_from_goal(goal))
+
+        return goals
+
+    async def _send_goals_to_semantic_belief_extractor(self, program: Program):
+        """
+        Extract goals from the program and send them to the Semantic Belief Extractor Agent.
+
+        :param program: The program received from the API.
+        """
+        goals = GoalList(goals=self._extract_goals_from_program(program))
+
+        message = InternalMessage(
+            to=settings.agent_settings.text_belief_extractor_name,
+            sender=self.name,
+            body=goals.model_dump_json(),
+            thread="goals",
+        )
+
+        await self.send(message)
+
     async def _receive_programs(self):
         """
         Continuous loop that receives program updates from the HTTP endpoint.
@@ -117,6 +156,7 @@ class BDIProgramManager(BaseAgent):
             await asyncio.gather(
                 self._create_agentspeak_and_send_to_bdi(program),
                 self._send_beliefs_to_semantic_belief_extractor(program),
+                self._send_goals_to_semantic_belief_extractor(program),
             )
 
     async def setup(self):

src/control_backend/agents/bdi/belief_collector_agent.py

@@ -101,7 +101,7 @@ class BDIBeliefCollectorAgent(BaseAgent):
             :return: A Belief object if the input is valid or None.
             """
             try:
-                return Belief(name=name, arguments=arguments, replace=name == "user_said")
+                return Belief(name=name, arguments=arguments)
             except ValidationError:
                 return None
 

src/control_backend/agents/bdi/text_belief_extractor_agent.py

@@ -2,17 +2,45 @@ import asyncio
 import json
 
 import httpx
-from pydantic import ValidationError
+from pydantic import BaseModel, ValidationError
 
 from control_backend.agents.base import BaseAgent
 from control_backend.agents.bdi.agentspeak_generator import AgentSpeakGenerator
 from control_backend.core.agent_system import InternalMessage
 from control_backend.core.config import settings
-from control_backend.schemas.belief_list import BeliefList
+from control_backend.schemas.belief_list import BeliefList, GoalList
 from control_backend.schemas.belief_message import Belief as InternalBelief
 from control_backend.schemas.belief_message import BeliefMessage
 from control_backend.schemas.chat_history import ChatHistory, ChatMessage
-from control_backend.schemas.program import SemanticBelief
+from control_backend.schemas.program import Goal, SemanticBelief
+
+type JSONLike = None | bool | int | float | str | list["JSONLike"] | dict[str, "JSONLike"]
+
+
+class BeliefState(BaseModel):
+    true: set[InternalBelief] = set()
+    false: set[InternalBelief] = set()
+
+    def difference(self, other: "BeliefState") -> "BeliefState":
+        return BeliefState(
+            true=self.true - other.true,
+            false=self.false - other.false,
+        )
+
+    def union(self, other: "BeliefState") -> "BeliefState":
+        return BeliefState(
+            true=self.true | other.true,
+            false=self.false | other.false,
+        )
+
+    def __sub__(self, other):
+        return self.difference(other)
+
+    def __or__(self, other):
+        return self.union(other)
+
+    def __bool__(self):
+        return bool(self.true) or bool(self.false)
 
 
 class TextBeliefExtractorAgent(BaseAgent):
@@ -27,12 +55,14 @@ class TextBeliefExtractorAgent(BaseAgent):
     the message itself.
     """
 
-    def __init__(self, name: str, temperature: float = settings.llm_settings.code_temperature):
+    def __init__(self, name: str):
         super().__init__(name)
-        self.beliefs: dict[str, bool] = {}
-        self.available_beliefs: list[SemanticBelief] = []
+        self._llm = self.LLM(self, settings.llm_settings.n_parallel)
+        self.belief_inferrer = SemanticBeliefInferrer(self._llm)
+        self.goal_inferrer = GoalAchievementInferrer(self._llm)
+        self._current_beliefs = BeliefState()
+        self._current_goal_completions: dict[str, bool] = {}
         self.conversation = ChatHistory(messages=[])
-        self.temperature = temperature
 
     async def setup(self):
         """
@@ -53,8 +83,9 @@ class TextBeliefExtractorAgent(BaseAgent):
             case settings.agent_settings.transcription_name:
                 self.logger.debug("Received text from transcriber: %s", msg.body)
                 self._apply_conversation_message(ChatMessage(role="user", content=msg.body))
-                await self._infer_new_beliefs()
                 await self._user_said(msg.body)
+                await self._infer_new_beliefs()
+                await self._infer_goal_completions()
             case settings.agent_settings.llm_name:
                 self.logger.debug("Received text from LLM: %s", msg.body)
                 self._apply_conversation_message(ChatMessage(role="assistant", content=msg.body))
@@ -76,10 +107,29 @@ class TextBeliefExtractorAgent(BaseAgent):
 
     def _handle_program_manager_message(self, msg: InternalMessage):
         """
-        Handle a message from the program manager: extract available beliefs from it.
+        Handle a message from the program manager: extract available beliefs and goals from it.
 
         :param msg: The received message from the program manager.
         """
+        match msg.thread:
+            case "beliefs":
+                self._handle_beliefs_message(msg)
+            case "goals":
+                self._handle_goals_message(msg)
+            case "conversation_history":
+                if msg.body == "reset":
+                    self._reset()
+            case _:
+                self.logger.warning("Received unexpected message from %s", msg.sender)
+
+    def _reset(self):
+        self.conversation = ChatHistory(messages=[])
+        self.belief_inferrer.available_beliefs.clear()
+        self._current_beliefs = BeliefState()
+        self.goal_inferrer.goals.clear()
+        self._current_goal_completions = {}
+
+    def _handle_beliefs_message(self, msg: InternalMessage):
         try:
             belief_list = BeliefList.model_validate_json(msg.body)
         except ValidationError:
@@ -88,10 +138,28 @@ class TextBeliefExtractorAgent(BaseAgent):
             )
             return
 
-        self.available_beliefs = [b for b in belief_list.beliefs if isinstance(b, SemanticBelief)]
+        available_beliefs = [b for b in belief_list.beliefs if isinstance(b, SemanticBelief)]
+        self.belief_inferrer.available_beliefs = available_beliefs
         self.logger.debug(
-            "Received %d beliefs from the program manager.",
-            len(self.available_beliefs),
+            "Received %d semantic beliefs from the program manager.",
+            len(available_beliefs),
+        )
+
+    def _handle_goals_message(self, msg: InternalMessage):
+        try:
+            goals_list = GoalList.model_validate_json(msg.body)
+        except ValidationError:
+            self.logger.warning(
+                "Received message from program manager but it is not a valid list of goals."
+            )
+            return
+
+        # Use only goals that can fail, as the others are always assumed to be completed
+        available_goals = [g for g in goals_list.goals if g.can_fail]
+        self.goal_inferrer.goals = available_goals
+        self.logger.debug(
+            "Received %d failable goals from the program manager.",
+            len(available_goals),
         )
 
     async def _user_said(self, text: str):
@@ -100,121 +168,210 @@ class TextBeliefExtractorAgent(BaseAgent):
 
         :param text: User's transcribed text.
         """
-        belief = {"beliefs": {"user_said": [text]}, "type": "belief_extraction_text"}
-        payload = json.dumps(belief)
-
         belief_msg = InternalMessage(
-            to=settings.agent_settings.bdi_belief_collector_name,
+            to=settings.agent_settings.bdi_core_name,
             sender=self.name,
-            body=payload,
+            body=BeliefMessage(
+                replace=[InternalBelief(name="user_said", arguments=[text])],
+            ).model_dump_json(),
             thread="beliefs",
         )
         await self.send(belief_msg)
 
     async def _infer_new_beliefs(self):
-        """
-        Process conversation history to extract beliefs, semantically. Any changed beliefs are sent
-        to the BDI core.
-        """
-        # Return instantly if there are no beliefs to infer
-        if not self.available_beliefs:
+        conversation_beliefs = await self.belief_inferrer.infer_from_conversation(self.conversation)
+
+        new_beliefs = conversation_beliefs - self._current_beliefs
+        if not new_beliefs:
             return
 
-        candidate_beliefs = await self._infer_turn()
-        belief_changes = BeliefMessage()
-        for belief_key, belief_value in candidate_beliefs.items():
-            if belief_value is None:
-                continue
-            old_belief_value = self.beliefs.get(belief_key)
-            if belief_value == old_belief_value:
-                continue
+        self._current_beliefs |= new_beliefs
 
-            self.beliefs[belief_key] = belief_value
+        belief_changes = BeliefMessage(
+            create=list(new_beliefs.true),
+            delete=list(new_beliefs.false),
+        )
 
-            belief = InternalBelief(name=belief_key, arguments=None)
-            if belief_value:
-                belief_changes.create.append(belief)
-            else:
-                belief_changes.delete.append(belief)
-
-        # Return if there were no changes in beliefs
-        if not belief_changes.has_values():
-            return
-
-        beliefs_message = InternalMessage(
+        message = InternalMessage(
             to=settings.agent_settings.bdi_core_name,
             sender=self.name,
             body=belief_changes.model_dump_json(),
             thread="beliefs",
         )
-        await self.send(beliefs_message)
+        await self.send(message)
 
-    @staticmethod
-    def _split_into_chunks[T](items: list[T], n: int) -> list[list[T]]:
-        k, m = divmod(len(items), n)
-        return [items[i * k + min(i, m) : (i + 1) * k + min(i + 1, m)] for i in range(n)]
+    async def _infer_goal_completions(self):
+        goal_completions = await self.goal_inferrer.infer_from_conversation(self.conversation)
 
-    async def _infer_turn(self) -> dict:
+        new_achieved = [
+            InternalBelief(name=goal, arguments=None)
+            for goal, achieved in goal_completions.items()
+            if achieved and self._current_goal_completions.get(goal) != achieved
+        ]
+        new_not_achieved = [
+            InternalBelief(name=goal, arguments=None)
+            for goal, achieved in goal_completions.items()
+            if not achieved and self._current_goal_completions.get(goal) != achieved
+        ]
+        for goal, achieved in goal_completions.items():
+            self._current_goal_completions[goal] = achieved
+
+        if not new_achieved and not new_not_achieved:
+            return
+
+        belief_changes = BeliefMessage(
+            create=new_achieved,
+            delete=new_not_achieved,
+        )
+        message = InternalMessage(
+            to=settings.agent_settings.bdi_core_name,
+            sender=self.name,
+            body=belief_changes.model_dump_json(),
+            thread="beliefs",
+        )
+        await self.send(message)
+
+    class LLM:
         """
-        Process the stored conversation history to extract semantic beliefs. Returns a list of
-        beliefs that have been set to ``True``, ``False`` or ``None``.
-
-        :return: A dict mapping belief names to a value ``True``, ``False`` or ``None``.
+        Class that handles sending structured generation requests to an LLM.
         """
+
+        def __init__(self, agent: "TextBeliefExtractorAgent", n_parallel: int):
+            self._agent = agent
+            self._semaphore = asyncio.Semaphore(n_parallel)
+
+        async def query(self, prompt: str, schema: dict, tries: int = 3) -> JSONLike | None:
+            """
+            Query the LLM with the given prompt and schema, return an instance of a dict conforming
+            to this schema. Try ``tries`` times, or return None.
+
+            :param prompt: Prompt to be queried.
+            :param schema: Schema to be queried.
+            :param tries: Number of times to try to query the LLM.
+            :return: An instance of a dict conforming to this schema, or None if failed.
+            """
+            try_count = 0
+            while try_count < tries:
+                try_count += 1
+
+                try:
+                    return await self._query_llm(prompt, schema)
+                except (httpx.HTTPError, json.JSONDecodeError, KeyError) as e:
+                    if try_count < tries:
+                        continue
+                    self._agent.logger.exception(
+                        "Failed to get LLM response after %d tries.",
+                        try_count,
+                        exc_info=e,
+                    )
+
+            return None
+
+        async def _query_llm(self, prompt: str, schema: dict) -> JSONLike:
+            """
+            Query an LLM with the given prompt and schema, return an instance of a dict conforming
+            to that schema.
+
+            :param prompt: The prompt to be queried.
+            :param schema: Schema to use during response.
+            :return: A dict conforming to this schema.
+            :raises httpx.HTTPStatusError: If the LLM server responded with an error.
+            :raises json.JSONDecodeError: If the LLM response was not valid JSON. May happen if the
+                response was cut off early due to length limitations.
+            :raises KeyError: If the LLM server responded with no error, but the response was
+                invalid.
+            """
+            async with self._semaphore:
+                async with httpx.AsyncClient() as client:
+                    response = await client.post(
+                        settings.llm_settings.local_llm_url,
+                        json={
+                            "model": settings.llm_settings.local_llm_model,
+                            "messages": [{"role": "user", "content": prompt}],
+                            "response_format": {
+                                "type": "json_schema",
+                                "json_schema": {
+                                    "name": "Beliefs",
+                                    "strict": True,
+                                    "schema": schema,
+                                },
+                            },
+                            "reasoning_effort": "low",
+                            "temperature": settings.llm_settings.code_temperature,
+                            "stream": False,
+                        },
+                        timeout=30.0,
+                    )
+                    response.raise_for_status()
+
+                    response_json = response.json()
+                    json_message = response_json["choices"][0]["message"]["content"]
+                    return json.loads(json_message)
+
+
+class SemanticBeliefInferrer:
+    """
+    Class that handles only prompting an LLM for semantic beliefs.
+    """
+
+    def __init__(
+        self,
+        llm: "TextBeliefExtractorAgent.LLM",
+        available_beliefs: list[SemanticBelief] | None = None,
+    ):
+        self._llm = llm
+        self.available_beliefs: list[SemanticBelief] = available_beliefs or []
+
+    async def infer_from_conversation(self, conversation: ChatHistory) -> BeliefState:
+        """
+        Process conversation history to extract beliefs, semantically. The result is an object that
+        describes all beliefs that hold or don't hold based on the full conversation.
+
+        :param conversation: The conversation history to be processed.
+        :return: An object that describes beliefs.
+        """
+        # Return instantly if there are no beliefs to infer
+        if not self.available_beliefs:
+            return BeliefState()
+
         n_parallel = max(1, min(settings.llm_settings.n_parallel - 1, len(self.available_beliefs)))
-        all_beliefs = await asyncio.gather(
+        all_beliefs: list[dict[str, bool | None] | None] = await asyncio.gather(
             *[
-                self._infer_beliefs(self.conversation, beliefs)
+                self._infer_beliefs(conversation, beliefs)
                 for beliefs in self._split_into_chunks(self.available_beliefs, n_parallel)
             ]
         )
-        retval = {}
+        retval = BeliefState()
         for beliefs in all_beliefs:
             if beliefs is None:
                 continue
-            retval.update(beliefs)
+            for belief_name, belief_holds in beliefs.items():
+                if belief_holds is None:
+                    continue
+                belief = InternalBelief(name=belief_name, arguments=None)
+                if belief_holds:
+                    retval.true.add(belief)
+                else:
+                    retval.false.add(belief)
         return retval
 
     @staticmethod
-    def _create_belief_schema(belief: SemanticBelief) -> tuple[str, dict]:
-        return AgentSpeakGenerator.slugify(belief), {
-            "type": ["boolean", "null"],
-            "description": belief.description,
-        }
+    def _split_into_chunks[T](items: list[T], n: int) -> list[list[T]]:
+        """
+        Split a list into ``n`` chunks, making each chunk approximately ``len(items) / n`` long.
 
-    @staticmethod
-    def _create_beliefs_schema(beliefs: list[SemanticBelief]) -> dict:
-        belief_schemas = [
-            TextBeliefExtractorAgent._create_belief_schema(belief) for belief in beliefs
-        ]
-
-        return {
-            "type": "object",
-            "properties": dict(belief_schemas),
-            "required": [name for name, _ in belief_schemas],
-        }
-
-    @staticmethod
-    def _format_message(message: ChatMessage):
-        return f"{message.role.upper()}:\n{message.content}"
-
-    @staticmethod
-    def _format_conversation(conversation: ChatHistory):
-        return "\n\n".join(
-            [TextBeliefExtractorAgent._format_message(message) for message in conversation.messages]
-        )
-
-    @staticmethod
-    def _format_beliefs(beliefs: list[SemanticBelief]):
-        return "\n".join(
-            [f"- {AgentSpeakGenerator.slugify(belief)}: {belief.description}" for belief in beliefs]
-        )
+        :param items: The list of items to split.
+        :param n: The number of desired chunks.
+        :return: A list of chunks each approximately ``len(items) / n`` long.
+        """
+        k, m = divmod(len(items), n)
+        return [items[i * k + min(i, m) : (i + 1) * k + min(i + 1, m)] for i in range(n)]
 
     async def _infer_beliefs(
         self,
         conversation: ChatHistory,
         beliefs: list[SemanticBelief],
-    ) -> dict | None:
+    ) -> dict[str, bool | None] | None:
         """
         Infer given beliefs based on the given conversation.
         :param conversation: The conversation to infer beliefs from.
@@ -241,70 +398,79 @@ Respond with a JSON similar to the following, but with the property names as giv
 
         schema = self._create_beliefs_schema(beliefs)
 
-        return await self._retry_query_llm(prompt, schema)
+        return await self._llm.query(prompt, schema)
 
-    async def _retry_query_llm(self, prompt: str, schema: dict, tries: int = 3) -> dict | None:
+    @staticmethod
+    def _create_belief_schema(belief: SemanticBelief) -> tuple[str, dict]:
+        return AgentSpeakGenerator.slugify(belief), {
+            "type": ["boolean", "null"],
+            "description": belief.description,
+        }
+
+    @staticmethod
+    def _create_beliefs_schema(beliefs: list[SemanticBelief]) -> dict:
+        belief_schemas = [
+            SemanticBeliefInferrer._create_belief_schema(belief) for belief in beliefs
+        ]
+
+        return {
+            "type": "object",
+            "properties": dict(belief_schemas),
+            "required": [name for name, _ in belief_schemas],
+        }
+
+    @staticmethod
+    def _format_message(message: ChatMessage):
+        return f"{message.role.upper()}:\n{message.content}"
+
+    @staticmethod
+    def _format_conversation(conversation: ChatHistory):
+        return "\n\n".join(
+            [SemanticBeliefInferrer._format_message(message) for message in conversation.messages]
+        )
+
+    @staticmethod
+    def _format_beliefs(beliefs: list[SemanticBelief]):
+        return "\n".join(
+            [f"- {AgentSpeakGenerator.slugify(belief)}: {belief.description}" for belief in beliefs]
+        )
+
+
+class GoalAchievementInferrer(SemanticBeliefInferrer):
+    def __init__(self, llm: TextBeliefExtractorAgent.LLM):
+        super().__init__(llm)
+        self.goals = []
+
+    async def infer_from_conversation(self, conversation: ChatHistory) -> dict[str, bool]:
         """
-        Query the LLM with the given prompt and schema, return an instance of a dict conforming
-        to this schema. Try ``tries`` times, or return None.
+        Determine which goals have been achieved based on the given conversation.
 
-        :param prompt: Prompt to be queried.
-        :param schema: Schema to be queried.
-        :return: An instance of a dict conforming to this schema, or None if failed.
+        :param conversation: The conversation to infer goal completion from.
+        :return: A mapping of goals and a boolean whether they have been achieved.
         """
-        try_count = 0
-        while try_count < tries:
-            try_count += 1
+        if not self.goals:
+            return {}
 
-            try:
-                return await self._query_llm(prompt, schema)
-            except (httpx.HTTPError, json.JSONDecodeError, KeyError) as e:
-                if try_count < tries:
-                    continue
-                self.logger.exception(
-                    "Failed to get LLM response after %d tries.",
-                    try_count,
-                    exc_info=e,
-                )
+        goals_achieved = await asyncio.gather(
+            *[self._infer_goal(conversation, g) for g in self.goals]
+        )
+        return {
+            f"achieved_{AgentSpeakGenerator.slugify(goal)}": achieved
+            for goal, achieved in zip(self.goals, goals_achieved, strict=True)
+        }
 
-        return None
+    async def _infer_goal(self, conversation: ChatHistory, goal: Goal) -> bool:
+        prompt = f"""{self._format_conversation(conversation)}
 
-    async def _query_llm(self, prompt: str, schema: dict) -> dict:
-        """
-        Query an LLM with the given prompt and schema, return an instance of a dict conforming to
-        that schema.
+Given the above conversation, what has the following goal been achieved?
 
-        :param prompt: The prompt to be queried.
-        :param schema: Schema to use during response.
-        :return: A dict conforming to this schema.
-        :raises httpx.HTTPStatusError: If the LLM server responded with an error.
-        :raises json.JSONDecodeError: If the LLM response was not valid JSON. May happen if the
-            response was cut off early due to length limitations.
-        :raises KeyError: If the LLM server responded with no error, but the response was invalid.
-        """
-        async with httpx.AsyncClient() as client:
-            response = await client.post(
-                settings.llm_settings.local_llm_url,
-                json={
-                    "model": settings.llm_settings.local_llm_model,
-                    "messages": [{"role": "user", "content": prompt}],
-                    "response_format": {
-                        "type": "json_schema",
-                        "json_schema": {
-                            "name": "Beliefs",
-                            "strict": True,
-                            "schema": schema,
-                        },
-                    },
-                    "reasoning_effort": "low",
-                    "temperature": self.temperature,
-                    "stream": False,
-                },
-                timeout=None,
-            )
-            response.raise_for_status()
+The name of the goal: {goal.name}
+Description of the goal: {goal.description}
 
-            response_json = response.json()
-            json_message = response_json["choices"][0]["message"]["content"]
-            beliefs = json.loads(json_message)
-            return beliefs
+Answer with literally only `true` or `false` (without backticks)."""
+
+        schema = {
+            "type": "boolean",
+        }
+
+        return await self._llm.query(prompt, schema)

src/control_backend/core/agent_system.py

@@ -192,7 +192,16 @@ class BaseAgent(ABC):
 
         :param coro: The coroutine to execute as a task.
         """
-        task = asyncio.create_task(coro)
+
+        async def try_coro(coro_: Coroutine):
+            try:
+                await coro_
+            except asyncio.CancelledError:
+                self.logger.debug("A behavior was canceled successfully: %s", coro_)
+            except Exception:
+                self.logger.warning("An exception occurred in a behavior.", exc_info=True)
+
+        task = asyncio.create_task(try_coro(coro))
         self._tasks.add(task)
         task.add_done_callback(self._tasks.discard)
         return task

src/control_backend/schemas/belief_list.py

@@ -1,6 +1,7 @@
 from pydantic import BaseModel
 
 from control_backend.schemas.program import Belief as ProgramBelief
+from control_backend.schemas.program import Goal
 
 
 class BeliefList(BaseModel):
@@ -12,3 +13,7 @@ class BeliefList(BaseModel):
     """
 
     beliefs: list[ProgramBelief]
+
+
+class GoalList(BaseModel):
+    goals: list[Goal]

src/control_backend/schemas/belief_message.py

@@ -13,6 +13,9 @@ class Belief(BaseModel):
     name: str
     arguments: list[str] | None
 
+    # To make it hashable
+    model_config = {"frozen": True}
+
 
 class BeliefMessage(BaseModel):
     """

src/control_backend/schemas/program.py

@@ -117,7 +117,7 @@ class Goal(ProgramElement):
     :ivar can_fail: Whether we can fail to achieve the goal after executing the plan.
     """
 
-    description: str
+    description: str = ""
     plan: Plan
     can_fail: bool = True