fix: some bugs and some tests

ref: N25B-429

.env.example

@@ -0,0 +1,20 @@
+# Example .env file. To use, make a copy, call it ".env" (i.e. removing the ".example" suffix), then you edit values.
+
+# The hostname of the Robot Interface. Change if the Control Backend and Robot Interface are running on different computers.
+RI_HOST="localhost"
+
+# URL for the local LLM API. Must be an API that implements the OpenAI Chat Completions API, but most do.
+LLM_SETTINGS__LOCAL_LLM_URL="http://localhost:1234/v1/chat/completions"
+
+# Name of the local LLM model to use.
+LLM_SETTINGS__LOCAL_LLM_MODEL="gpt-oss"
+
+# Number of non-speech chunks to wait before speech ended. A chunk is approximately 31 ms. Increasing this number allows longer pauses in speech, but also increases response time.
+BEHAVIOUR_SETTINGS__VAD_NON_SPEECH_PATIENCE_CHUNKS=15
+
+# Timeout in milliseconds for socket polling. Increase this number if network latency/jitter is high, often the case when using Wi-Fi. Perhaps 500 ms. A symptom of this issue is transcriptions getting cut off.
+BEHAVIOUR_SETTINGS__SOCKET_POLLER_TIMEOUT_MS=100
+
+
+
+# For an exhaustive list of options, see the control_backend.core.config module in the docs.

README.md

@@ -27,6 +27,7 @@ This + part might differ based on what model you choose.
 copy the model name in the module loaded and replace local_llm_modelL. In settings.
 
 
+
 ## Running
 To run the project (development server), execute the following command (while inside the root repository):
 
@@ -34,6 +35,14 @@ To run the project (development server), execute the following command (while in
 uv run fastapi dev src/control_backend/main.py
 ```
 
+### Environment Variables
+
+You can use environment variables to change settings. Make a copy of the [`.env.example`](.env.example) file, name it `.env` and put it in the root directory. The file itself describes how to do the configuration.
+
+For an exhaustive list of environment options, see the `control_backend.core.config` module in the docs.
+
+
+
 ## Testing
 Testing happens automatically when opening a merge request to any branch. If you want to manually run the test suite, you can do so by running the following for unit tests:
 

src/control_backend/agents/actuation/robot_gesture_agent.py

@@ -33,7 +33,7 @@ class RobotGestureAgent(BaseAgent):
     def __init__(
         self,
         name: str,
-        address=settings.zmq_settings.ri_command_address,
+        address: str,
         bind=False,
         gesture_data=None,
         single_gesture_data=None,

src/control_backend/agents/bdi/agentspeak_generator.py

@@ -145,7 +145,10 @@ class AgentSpeakGenerator:
                 type=TriggerType.ADDED_BELIEF,
                 trigger_literal=AstLiteral("user_said", [AstVar("Message")]),
                 context=[AstLiteral("phase", [AstString("end")])],
-                body=[AstStatement(StatementType.ACHIEVE_GOAL, AstLiteral("reply"))],
+                body=[
+                    AstStatement(StatementType.DO_ACTION, AstLiteral("notify_user_said")),
+                    AstStatement(StatementType.ACHIEVE_GOAL, AstLiteral("reply")),
+                ],
             )
         )
 
@@ -157,7 +160,7 @@ class AgentSpeakGenerator:
 
         previous_goal = None
         for goal in phase.goals:
-            self._process_goal(goal, phase, previous_goal)
+            self._process_goal(goal, phase, previous_goal, main_goal=True)
             previous_goal = goal
 
         for trigger in phase.triggers:
@@ -171,26 +174,41 @@ class AgentSpeakGenerator:
             self._astify(to_phase) if to_phase else AstLiteral("phase", [AstString("end")])
         )
 
-        context = [from_phase_ast, ~AstLiteral("responded_this_turn")]
-        if from_phase and from_phase.goals:
-            context.append(self._astify(from_phase.goals[-1], achieved=True))
+        context = [from_phase_ast]
+        if from_phase:
+            for goal in from_phase.goals:
+                context.append(self._astify(goal, achieved=True))
 
         body = [
             AstStatement(StatementType.REMOVE_BELIEF, from_phase_ast),
             AstStatement(StatementType.ADD_BELIEF, to_phase_ast),
         ]
 
-        if from_phase:
-            body.extend(
-                [
-                    AstStatement(
-                        StatementType.TEST_GOAL, AstLiteral("user_said", [AstVar("Message")])
-                    ),
-                    AstStatement(
-                        StatementType.REPLACE_BELIEF, AstLiteral("user_said", [AstVar("Message")])
-                    ),
-                ]
+        # if from_phase:
+        #     body.extend(
+        #         [
+        #             AstStatement(
+        #                 StatementType.TEST_GOAL, AstLiteral("user_said", [AstVar("Message")])
+        #             ),
+        #             AstStatement(
+        #                 StatementType.REPLACE_BELIEF, AstLiteral("user_said", [AstVar("Message")])
+        #             ),
+        #         ]
+        #     )
+
+        # Notify outside world about transition
+        body.append(
+            AstStatement(
+                StatementType.DO_ACTION,
+                AstLiteral(
+                    "notify_transition_phase",
+                    [
+                        AstString(str(from_phase.id)),
+                        AstString(str(to_phase.id) if to_phase else "end"),
+                    ],
+                ),
             )
+        )
 
         self._asp.plans.append(
             AstPlan(TriggerType.ADDED_GOAL, AstLiteral("transition_phase"), context, body)
@@ -213,6 +231,11 @@ class AgentSpeakGenerator:
     def _add_default_loop(self, phase: Phase) -> None:
         actions = []
 
+        actions.append(
+            AstStatement(
+                StatementType.DO_ACTION, AstLiteral("notify_user_said", [AstVar("Message")])
+            )
+        )
         actions.append(AstStatement(StatementType.REMOVE_BELIEF, AstLiteral("responded_this_turn")))
         actions.append(AstStatement(StatementType.ACHIEVE_GOAL, AstLiteral("check_triggers")))
 
@@ -236,6 +259,7 @@ class AgentSpeakGenerator:
         phase: Phase,
         previous_goal: Goal | None = None,
         continues_response: bool = False,
+        main_goal: bool = False,
     ) -> None:
         context: list[AstExpression] = [self._astify(phase)]
         context.append(~self._astify(goal, achieved=True))
@@ -245,6 +269,13 @@ class AgentSpeakGenerator:
             context.append(~AstLiteral("responded_this_turn"))
 
         body = []
+        if main_goal:  # UI only needs to know about the main goals
+            body.append(
+                AstStatement(
+                    StatementType.DO_ACTION,
+                    AstLiteral("notify_goal_start", [AstString(self.slugify(goal))]),
+                )
+            )
 
         subgoals = []
         for step in goal.plan.steps:
@@ -283,11 +314,23 @@ class AgentSpeakGenerator:
         body = []
         subgoals = []
 
+        body.append(
+            AstStatement(
+                StatementType.DO_ACTION,
+                AstLiteral("notify_trigger_start", [AstString(self.slugify(trigger))]),
+            )
+        )
         for step in trigger.plan.steps:
             body.append(self._step_to_statement(step))
             if isinstance(step, Goal):
                 step.can_fail = False  # triggers are continuous sequence
                 subgoals.append(step)
+        body.append(
+            AstStatement(
+                StatementType.DO_ACTION,
+                AstLiteral("notify_trigger_end", [AstString(self.slugify(trigger))]),
+            )
+        )
 
         self._asp.plans.append(
             AstPlan(
@@ -298,6 +341,9 @@ class AgentSpeakGenerator:
             )
         )
 
+        # Force trigger (from UI)
+        self._asp.plans.append(AstPlan(TriggerType.ADDED_GOAL, self._astify(trigger), [], body))
+
         for subgoal in subgoals:
             self._process_goal(subgoal, phase, continues_response=True)
 
@@ -332,13 +378,7 @@ class AgentSpeakGenerator:
 
     @_astify.register
     def _(self, sb: SemanticBelief) -> AstExpression:
-        return AstLiteral(self.get_semantic_belief_slug(sb))
-
-    @staticmethod
-    def get_semantic_belief_slug(sb: SemanticBelief) -> str:
-        # If you need a method like this for other types, make a public slugify singledispatch for
-        #  all types.
-        return f"semantic_{AgentSpeakGenerator._slugify_str(sb.name)}"
+        return AstLiteral(self.slugify(sb))
 
     @_astify.register
     def _(self, ib: InferredBelief) -> AstExpression:

src/control_backend/agents/bdi/bdi_core_agent.py

@@ -1,5 +1,6 @@
 import asyncio
 import copy
+import json
 import time
 from collections.abc import Iterable
 
@@ -13,7 +14,7 @@ from control_backend.core.agent_system import InternalMessage
 from control_backend.core.config import settings
 from control_backend.schemas.belief_message import BeliefMessage
 from control_backend.schemas.llm_prompt_message import LLMPromptMessage
-from control_backend.schemas.ri_message import SpeechCommand
+from control_backend.schemas.ri_message import GestureCommand, RIEndpoint, SpeechCommand
 
 DELIMITER = ";\n"  # TODO: temporary until we support lists in AgentSpeak
 
@@ -100,7 +101,6 @@ class BDICoreAgent(BaseAgent):
             maybe_more_work = True
             while maybe_more_work:
                 maybe_more_work = False
-                self.logger.debug("Stepping BDI.")
                 if self.bdi_agent.step():
                     maybe_more_work = True
 
@@ -155,6 +155,17 @@ class BDICoreAgent(BaseAgent):
                     body=cmd.model_dump_json(),
                 )
                 await self.send(out_msg)
+            case settings.agent_settings.user_interrupt_name:
+                content = msg.body
+                self.logger.debug("Received user interruption: %s", content)
+
+                match msg.thread:
+                    case "force_phase_transition":
+                        self._set_goal("transition_phase")
+                    case "force_trigger":
+                        self._force_trigger(msg.body)
+                    case _:
+                        self.logger.warning("Received unknow user interruption: %s", msg)
 
     def _apply_belief_changes(self, belief_changes: BeliefMessage):
         """
@@ -201,16 +212,35 @@ class BDICoreAgent(BaseAgent):
             agentspeak.runtime.Intention(),
         )
 
+        # Check for transitions
+        self.bdi_agent.call(
+            agentspeak.Trigger.addition,
+            agentspeak.GoalType.achievement,
+            agentspeak.Literal("transition_phase"),
+            agentspeak.runtime.Intention(),
+        )
+
+        # Check triggers
+        self.bdi_agent.call(
+            agentspeak.Trigger.addition,
+            agentspeak.GoalType.achievement,
+            agentspeak.Literal("check_triggers"),
+            agentspeak.runtime.Intention(),
+        )
+
         self._wake_bdi_loop.set()
 
         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,
@@ -250,6 +280,37 @@ class BDICoreAgent(BaseAgent):
 
         self.logger.debug(f"Removed {removed_count} beliefs.")
 
+    def _set_goal(self, name: str, args: Iterable[str] | None = None):
+        args = args or []
+
+        if args:
+            merged_args = DELIMITER.join(arg for arg in args)
+            new_args = (agentspeak.Literal(merged_args),)
+            term = agentspeak.Literal(name, new_args)
+        else:
+            term = agentspeak.Literal(name)
+
+        self.bdi_agent.call(
+            agentspeak.Trigger.addition,
+            agentspeak.GoalType.achievement,
+            term,
+            agentspeak.runtime.Intention(),
+        )
+
+        self._wake_bdi_loop.set()
+
+        self.logger.debug(f"Set goal !{self.format_belief_string(name, args)}.")
+
+    def _force_trigger(self, name: str):
+        self.bdi_agent.call(
+            agentspeak.Trigger.addition,
+            agentspeak.GoalType.achievement,
+            agentspeak.Literal(name),
+            agentspeak.runtime.Intention(),
+        )
+
+        self.logger.info("Manually forced trigger %s.", name)
+
     def _add_custom_actions(self) -> None:
         """
         Add any custom actions here. Inside `@self.actions.add()`, the first argument is
@@ -258,7 +319,7 @@ class BDICoreAgent(BaseAgent):
         """
 
         @self.actions.add(".reply", 2)
-        def _reply(agent: "BDICoreAgent", term, intention):
+        def _reply(agent, term, intention):
             """
             Let the LLM generate a response to a user's utterance with the current norms and goals.
             """
@@ -291,7 +352,7 @@ class BDICoreAgent(BaseAgent):
             yield
 
         @self.actions.add(".say", 1)
-        def _say(agent: "BDICoreAgent", term, intention):
+        def _say(agent, term, intention):
             """
             Make the robot say the given text instantly.
             """
@@ -305,12 +366,21 @@ class BDICoreAgent(BaseAgent):
                 sender=settings.agent_settings.bdi_core_name,
                 body=speech_command.model_dump_json(),
             )
-            # TODO: add to conversation history
+
             self.add_behavior(self.send(speech_message))
+
+            chat_history_message = InternalMessage(
+                to=settings.agent_settings.llm_name,
+                thread="assistant_message",
+                body=str(message_text),
+            )
+
+            self.add_behavior(self.send(chat_history_message))
+
             yield
 
         @self.actions.add(".gesture", 2)
-        def _gesture(agent: "BDICoreAgent", term, intention):
+        def _gesture(agent, term, intention):
             """
             Make the robot perform the given gesture instantly.
             """
@@ -323,13 +393,113 @@ class BDICoreAgent(BaseAgent):
                 gesture_name,
             )
 
-            # gesture = Gesture(type=gesture_type, name=gesture_name)
-            # gesture_message = InternalMessage(
-            #     to=settings.agent_settings.robot_gesture_name,
-            #     sender=settings.agent_settings.bdi_core_name,
-            #     body=gesture.model_dump_json(),
-            # )
-            # asyncio.create_task(agent.send(gesture_message))
+            if str(gesture_type) == "single":
+                endpoint = RIEndpoint.GESTURE_SINGLE
+            elif str(gesture_type) == "tag":
+                endpoint = RIEndpoint.GESTURE_TAG
+            else:
+                self.logger.warning("Gesture type %s could not be resolved.", gesture_type)
+                endpoint = RIEndpoint.GESTURE_SINGLE
+
+            gesture_command = GestureCommand(endpoint=endpoint, data=gesture_name)
+            gesture_message = InternalMessage(
+                to=settings.agent_settings.robot_gesture_name,
+                sender=settings.agent_settings.bdi_core_name,
+                body=gesture_command.model_dump_json(),
+            )
+            self.add_behavior(self.send(gesture_message))
+            yield
+
+        @self.actions.add(".notify_user_said", 1)
+        def _notify_user_said(agent, term, intention):
+            user_said = agentspeak.grounded(term.args[0], intention.scope)
+
+            msg = InternalMessage(
+                to=settings.agent_settings.llm_name, thread="user_message", body=str(user_said)
+            )
+
+            self.add_behavior(self.send(msg))
+
+            yield
+
+        @self.actions.add(".notify_trigger_start", 1)
+        def _notify_trigger_start(agent, term, intention):
+            """
+            Notify the UI about the trigger we just started doing.
+            """
+            trigger_name = agentspeak.grounded(term.args[0], intention.scope)
+
+            self.logger.debug("Started trigger %s", trigger_name)
+
+            msg = InternalMessage(
+                to=settings.agent_settings.user_interrupt_name,
+                sender=self.name,
+                thread="trigger_start",
+                body=str(trigger_name),
+            )
+
+            # TODO: check with Pim
+            self.add_behavior(self.send(msg))
+
+            yield
+
+        @self.actions.add(".notify_trigger_end", 1)
+        def _notify_trigger_end(agent, term, intention):
+            """
+            Notify the UI about the trigger we just started doing.
+            """
+            trigger_name = agentspeak.grounded(term.args[0], intention.scope)
+
+            self.logger.debug("Finished trigger %s", trigger_name)
+
+            msg = InternalMessage(
+                to=settings.agent_settings.user_interrupt_name,
+                sender=self.name,
+                thread="trigger_end",
+                body=str(trigger_name),
+            )
+
+            # TODO: check with Pim
+            self.add_behavior(self.send(msg))
+
+            yield
+
+        @self.actions.add(".notify_goal_start", 1)
+        def _notify_goal_start(agent, term, intention):
+            """
+            Notify the UI about the goal we just started chasing.
+            """
+            goal_name = agentspeak.grounded(term.args[0], intention.scope)
+
+            self.logger.debug("Started chasing goal %s", goal_name)
+
+            msg = InternalMessage(
+                to=settings.agent_settings.user_interrupt_name,
+                sender=self.name,
+                thread="goal_start",
+                body=str(goal_name),
+            )
+
+            self.add_behavior(self.send(msg))
+
+            yield
+
+        @self.actions.add(".notify_transition_phase", 2)
+        def _notify_transition_phase(agent, term, intention):
+            """
+            Notify the BDI program manager about a phase transition.
+            """
+            old = agentspeak.grounded(term.args[0], intention.scope)
+            new = agentspeak.grounded(term.args[1], intention.scope)
+
+            msg = InternalMessage(
+                to=settings.agent_settings.bdi_program_manager_name,
+                thread="transition_phase",
+                body=json.dumps({"old": str(old), "new": str(new)}),
+            )
+
+            self.add_behavior(self.send(msg))
+
             yield
 
     async def _send_to_llm(self, text: str, norms: str, goals: str):
@@ -341,13 +511,14 @@ class BDICoreAgent(BaseAgent):
             to=settings.agent_settings.llm_name,
             sender=self.name,
             body=prompt.model_dump_json(),
+            thread="prompt_message",
         )
         await self.send(msg)
         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

@@ -1,4 +1,5 @@
 import asyncio
+import json
 
 import zmq
 from pydantic import ValidationError
@@ -7,9 +8,16 @@ 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,
+    Phase,
+    Program,
+)
 
 
 class BDIProgramManager(BaseAgent):
@@ -24,20 +32,20 @@ class BDIProgramManager(BaseAgent):
     :ivar sub_socket: The ZMQ SUB socket used to receive program updates.
     """
 
+    _program: Program
+    _phase: Phase | None
+
     def __init__(self, **kwargs):
         super().__init__(**kwargs)
         self.sub_socket = None
 
+    def _initialize_internal_state(self, program: Program):
+        self._program = program
+        self._phase = program.phases[0]  # start in first phase
+
     async def _create_agentspeak_and_send_to_bdi(self, program: Program):
         """
-        Convert a received program into BDI beliefs and send them to the BDI Core Agent.
-
-        Currently, it takes the **first phase** of the program and extracts:
-        -   **Norms**: Constraints or rules the agent must follow.
-        -   **Goals**: Objectives the agent must achieve.
-
-        These are sent as a ``BeliefMessage`` with ``replace=True``, meaning they will
-        overwrite any existing norms/goals of the same name in the BDI agent.
+        Convert a received program into an AgentSpeak file and send it to the BDI Core Agent.
 
         :param program: The program object received from the API.
         """
@@ -59,17 +67,45 @@ class BDIProgramManager(BaseAgent):
 
         await self.send(msg)
 
-    @staticmethod
-    def _extract_beliefs_from_program(program: Program) -> list[Belief]:
+    async def handle_message(self, msg: InternalMessage):
+        match msg.thread:
+            case "transition_phase":
+                phases = json.loads(msg.body)
+
+                await self._transition_phase(phases["old"], phases["new"])
+
+    async def _transition_phase(self, old: str, new: str):
+        assert old == str(self._phase.id)
+
+        if new == "end":
+            self._phase = None
+            return
+
+        for phase in self._program.phases:
+            if str(phase.id) == new:
+                self._phase = phase
+
+        await self._send_beliefs_to_semantic_belief_extractor()
+        await self._send_goals_to_semantic_belief_extractor()
+
+        # Notify user interaction agent
+        msg = InternalMessage(
+            to=settings.agent_settings.user_interrupt_name,
+            thread="transition_phase",
+            body=str(self._phase.id),
+        )
+
+        self.add_behavior(self.send(msg))
+
+    def _extract_current_beliefs(self) -> list[Belief]:
         beliefs: list[Belief] = []
 
-        for phase in program.phases:
-            for norm in phase.norms:
-                if isinstance(norm, ConditionalNorm):
-                    beliefs += BDIProgramManager._extract_beliefs_from_belief(norm.condition)
+        for norm in self._phase.norms:
+            if isinstance(norm, ConditionalNorm):
+                beliefs += self._extract_beliefs_from_belief(norm.condition)
 
-            for trigger in phase.triggers:
-                beliefs += BDIProgramManager._extract_beliefs_from_belief(trigger.condition)
+        for trigger in self._phase.triggers:
+            beliefs += self._extract_beliefs_from_belief(trigger.condition)
 
         return beliefs
 
@@ -81,13 +117,11 @@ class BDIProgramManager(BaseAgent):
             ) + BDIProgramManager._extract_beliefs_from_belief(belief.right)
         return [belief]
 
-    async def _send_beliefs_to_semantic_belief_extractor(self, program: Program):
+    async def _send_beliefs_to_semantic_belief_extractor(self):
         """
         Extract beliefs from the program and send them to the Semantic Belief Extractor Agent.
-
-        :param program: The program received from the API.
         """
-        beliefs = BeliefList(beliefs=self._extract_beliefs_from_program(program))
+        beliefs = BeliefList(beliefs=self._extract_current_beliefs())
 
         message = InternalMessage(
             to=settings.agent_settings.text_belief_extractor_name,
@@ -98,12 +132,69 @@ class BDIProgramManager(BaseAgent):
 
         await self.send(message)
 
+    def _extract_current_goals(self) -> list[Goal]:
+        """
+        Extract all goals from the program, including subgoals.
+
+        :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 goal in self._phase.goals:
+            goals.extend(extract_goals_from_goal(goal))
+
+        return goals
+
+    async def _send_goals_to_semantic_belief_extractor(self):
+        """
+        Extract goals for the current phase and send them to the Semantic Belief Extractor Agent.
+        """
+        goals = GoalList(goals=self._extract_current_goals())
+
+        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 _send_clear_llm_history(self):
+        """
+        Clear the LLM Agent's conversation history.
+
+        Sends an empty history to the LLM Agent to reset its state.
+        """
+        message = InternalMessage(
+            to=settings.agent_settings.llm_name,
+            body="clear_history",
+        )
+        await self.send(message)
+        self.logger.debug("Sent message to LLM agent to clear history.")
+
+        extractor_msg = InternalMessage(
+            to=settings.agent_settings.text_belief_extractor_name,
+            thread="conversation_history",
+            body="reset",
+        )
+        await self.send(extractor_msg)
+        self.logger.debug("Sent message to extractor agent to clear history.")
+
     async def _receive_programs(self):
         """
         Continuous loop that receives program updates from the HTTP endpoint.
 
         It listens to the ``program`` topic on the internal ZMQ SUB socket.
         When a program is received, it is validated and forwarded to BDI via :meth:`_send_to_bdi`.
+        Additionally, the LLM history is cleared via :meth:`_send_clear_llm_history`.
         """
         while True:
             topic, body = await self.sub_socket.recv_multipart()
@@ -111,12 +202,17 @@ class BDIProgramManager(BaseAgent):
             try:
                 program = Program.model_validate_json(body)
             except ValidationError:
-                self.logger.exception("Received an invalid program.")
+                self.logger.warning("Received an invalid program.")
                 continue
 
+            self._initialize_internal_state(program)
+
+            await self._send_clear_llm_history()
+
             await asyncio.gather(
                 self._create_agentspeak_and_send_to_bdi(program),
-                self._send_beliefs_to_semantic_belief_extractor(program),
+                self._send_beliefs_to_semantic_belief_extractor(),
+                self._send_goals_to_semantic_belief_extractor(),
             )
 
     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/default_behavior.asl

@@ -1,5 +1,6 @@
 norms("").
 
 +user_said(Message) : norms(Norms) <-
+  .notify_user_said(Message);
   -user_said(Message);
   .reply(Message, Norms).

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):
         """
@@ -53,13 +83,14 @@ 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))
             case settings.agent_settings.bdi_program_manager_name:
-                self._handle_program_manager_message(msg)
+                await self._handle_program_manager_message(msg)
             case _:
                 self.logger.info("Discarding message from %s", sender)
                 return
@@ -74,12 +105,33 @@ class TextBeliefExtractorAgent(BaseAgent):
         length_limit = settings.behaviour_settings.conversation_history_length_limit
         self.conversation.messages = (self.conversation.messages + [message])[-length_limit:]
 
-    def _handle_program_manager_message(self, msg: InternalMessage):
+    async 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)
+                await self._infer_new_beliefs()
+            case "goals":
+                self._handle_goals_message(msg)
+                await self._infer_goal_completions()
+            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 +140,30 @@ 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: %s",
+            len(available_beliefs),
+            ", ".join(b.name for b in 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: %s",
+            len(available_goals),
+            ", ".join(g.name for g in available_goals),
         )
 
     async def _user_said(self, text: str):
@@ -100,121 +172,212 @@ 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:
+            self.logger.debug("No new beliefs detected.")
             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:
+            self.logger.debug("No goal achievement changes detected.")
+            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 +404,79 @@ 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/agents/communication/ri_communication_agent.py

@@ -38,7 +38,7 @@ class RICommunicationAgent(BaseAgent):
     def __init__(
         self,
         name: str,
-        address=settings.zmq_settings.ri_command_address,
+        address=settings.zmq_settings.ri_communication_address,
         bind=False,
     ):
         super().__init__(name)
@@ -168,7 +168,7 @@ class RICommunicationAgent(BaseAgent):
             bind = port_data["bind"]
 
             if not bind:
-                addr = f"tcp://localhost:{port}"
+                addr = f"tcp://{settings.ri_host}:{port}"
             else:
                 addr = f"tcp://*:{port}"
 
@@ -248,7 +248,8 @@ class RICommunicationAgent(BaseAgent):
                     self._req_socket.recv_json(), timeout=seconds_to_wait_total / 2
                 )
 
-                self.logger.debug(f'Received message "{message}" from RI.')
+                if "endpoint" in message and message["endpoint"] != "ping":
+                    self.logger.debug(f'Received message "{message}" from RI.')
                 if "endpoint" not in message:
                     self.logger.warning("No received endpoint in message, expected ping endpoint.")
                     continue

src/control_backend/agents/llm/llm_agent.py

@@ -46,14 +46,23 @@ class LLMAgent(BaseAgent):
         :param msg: The received internal message.
         """
         if msg.sender == settings.agent_settings.bdi_core_name:
-            self.logger.debug("Processing message from BDI core.")
-            try:
-                prompt_message = LLMPromptMessage.model_validate_json(msg.body)
-                await self._process_bdi_message(prompt_message)
-            except ValidationError:
-                self.logger.debug("Prompt message from BDI core is invalid.")
+            match msg.thread:
+                case "prompt_message":
+                    try:
+                        prompt_message = LLMPromptMessage.model_validate_json(msg.body)
+                        await self._process_bdi_message(prompt_message)
+                    except ValidationError:
+                        self.logger.debug("Prompt message from BDI core is invalid.")
+                case "assistant_message":
+                    self.history.append({"role": "assistant", "content": msg.body})
+                case "user_message":
+                    self.history.append({"role": "user", "content": msg.body})
+        elif msg.sender == settings.agent_settings.bdi_program_manager_name:
+            if msg.body == "clear_history":
+                self.logger.debug("Clearing conversation history.")
+                self.history.clear()
         else:
-            self.logger.debug("Message ignored (not from BDI core.")
+            self.logger.debug("Message ignored.")
 
     async def _process_bdi_message(self, message: LLMPromptMessage):
         """
@@ -114,13 +123,6 @@ class LLMAgent(BaseAgent):
         :param goals: Goals the LLM should achieve.
         :yield: Fragments of the LLM-generated content (e.g., sentences/phrases).
         """
-        self.history.append(
-            {
-                "role": "user",
-                "content": prompt,
-            }
-        )
-
         instructions = LLMInstructions(norms if norms else None, goals if goals else None)
         messages = [
             {

src/control_backend/agents/perception/vad_agent.py

@@ -103,12 +103,11 @@ class VADAgent(BaseAgent):
 
         self._connect_audio_in_socket()
 
-        audio_out_port = self._connect_audio_out_socket()
-        if audio_out_port is None:
+        audio_out_address = self._connect_audio_out_socket()
+        if audio_out_address is None:
             self.logger.error("Could not bind output socket, stopping.")
             await self.stop()
             return
-        audio_out_address = f"tcp://localhost:{audio_out_port}"
 
         # Connect to internal communication socket
         self.program_sub_socket = azmq.Context.instance().socket(zmq.SUB)
@@ -161,13 +160,14 @@ class VADAgent(BaseAgent):
             self.audio_in_socket.connect(self.audio_in_address)
         self.audio_in_poller = SocketPoller[bytes](self.audio_in_socket)
 
-    def _connect_audio_out_socket(self) -> int | None:
+    def _connect_audio_out_socket(self) -> str | None:
         """
-        Returns the port bound, or None if binding failed.
+        Returns the address that was bound to, or None if binding failed.
         """
         try:
             self.audio_out_socket = azmq.Context.instance().socket(zmq.PUB)
-            return self.audio_out_socket.bind_to_random_port("tcp://localhost", max_tries=100)
+            self.audio_out_socket.bind(settings.zmq_settings.vad_pub_address)
+            return settings.zmq_settings.vad_pub_address
         except zmq.ZMQBindError:
             self.logger.error("Failed to bind an audio output socket after 100 tries.")
             self.audio_out_socket = None
@@ -229,10 +229,11 @@ class VADAgent(BaseAgent):
             assert self.model is not None
             prob = self.model(torch.from_numpy(chunk), settings.vad_settings.sample_rate_hz).item()
             non_speech_patience = settings.behaviour_settings.vad_non_speech_patience_chunks
+            begin_silence_length = settings.behaviour_settings.vad_begin_silence_chunks
             prob_threshold = settings.behaviour_settings.vad_prob_threshold
 
             if prob > prob_threshold:
-                if self.i_since_speech > non_speech_patience:
+                if self.i_since_speech > non_speech_patience + begin_silence_length:
                     self.logger.debug("Speech started.")
                 self.audio_buffer = np.append(self.audio_buffer, chunk)
                 self.i_since_speech = 0
@@ -246,11 +247,12 @@ class VADAgent(BaseAgent):
                 continue
 
             # Speech probably ended. Make sure we have a usable amount of data.
-            if len(self.audio_buffer) >= 3 * len(chunk):
+            if len(self.audio_buffer) > begin_silence_length * len(chunk):
                 self.logger.debug("Speech ended.")
                 assert self.audio_out_socket is not None
                 await self.audio_out_socket.send(self.audio_buffer[: -2 * len(chunk)].tobytes())
 
-            # At this point, we know that the speech has ended.
-            # Prepend the last chunk that had no speech, for a more fluent boundary
-            self.audio_buffer = chunk
+            # At this point, we know that there is no speech.
+            # Prepend the last few chunks that had no speech, for a more fluent boundary.
+            self.audio_buffer = np.append(self.audio_buffer, chunk)
+            self.audio_buffer = self.audio_buffer[-begin_silence_length * len(chunk) :]

src/control_backend/core/agent_system.py

@@ -131,6 +131,7 @@ class BaseAgent(ABC):
         :param message: The message to send.
         """
         target = AgentDirectory.get(message.to)
+        message.sender = self.name
         if target:
             await target.inbox.put(message)
             self.logger.debug(f"Sent message {message.body} to {message.to} via regular inbox.")
@@ -192,7 +193,16 @@ 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/core/config.py

@@ -1,3 +1,12 @@
+"""
+An exhaustive overview of configurable options. All of these can be set using environment variables
+by nesting with double underscores (__). Start from the ``Settings`` class.
+
+For example, ``settings.ri_host`` becomes ``RI_HOST``, and
+``settings.zmq_settings.ri_communication_address`` becomes
+``ZMQ_SETTINGS__RI_COMMUNICATION_ADDRESS``.
+"""
+
 from pydantic import BaseModel
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
@@ -8,16 +17,17 @@ class ZMQSettings(BaseModel):
 
     :ivar internal_pub_address: Address for the internal PUB socket.
     :ivar internal_sub_address: Address for the internal SUB socket.
-    :ivar ri_command_address: Address for sending commands to the Robot Interface.
-    :ivar ri_communication_address: Address for receiving communication from the Robot Interface.
-    :ivar vad_agent_address: Address for the Voice Activity Detection (VAD) agent.
+    :ivar ri_communication_address: Address for the endpoint that the Robot Interface connects to.
+    :ivar vad_pub_address: Address that the VAD agent binds to and publishes audio segments to.
     """
 
+    # ATTENTION: When adding/removing settings, make sure to update the .env.example file
+
     internal_pub_address: str = "tcp://localhost:5560"
     internal_sub_address: str = "tcp://localhost:5561"
-    ri_command_address: str = "tcp://localhost:0000"
     ri_communication_address: str = "tcp://*:5555"
     internal_gesture_rep_adress: str = "tcp://localhost:7788"
+    vad_pub_address: str = "inproc://vad_stream"
 
 
 class AgentSettings(BaseModel):
@@ -36,6 +46,8 @@ class AgentSettings(BaseModel):
     :ivar robot_speech_name: Name of the Robot Speech Agent.
     """
 
+    # ATTENTION: When adding/removing settings, make sure to update the .env.example file
+
     # agent names
     bdi_core_name: str = "bdi_core_agent"
     bdi_belief_collector_name: str = "belief_collector_agent"
@@ -61,6 +73,7 @@ class BehaviourSettings(BaseModel):
     :ivar vad_prob_threshold: Probability threshold for Voice Activity Detection.
     :ivar vad_initial_since_speech: Initial value for 'since speech' counter in VAD.
     :ivar vad_non_speech_patience_chunks: Number of non-speech chunks to wait before speech ended.
+    :ivar vad_begin_silence_chunks: The number of chunks of silence to prepend to speech chunks.
     :ivar transcription_max_concurrent_tasks: Maximum number of concurrent transcription tasks.
     :ivar transcription_words_per_minute: Estimated words per minute for transcription timing.
     :ivar transcription_words_per_token: Estimated words per token for transcription timing.
@@ -68,6 +81,8 @@ class BehaviourSettings(BaseModel):
     :ivar conversation_history_length_limit: The maximum amount of messages to extract beliefs from.
     """
 
+    # ATTENTION: When adding/removing settings, make sure to update the .env.example file
+
     sleep_s: float = 1.0
     comm_setup_max_retries: int = 5
     socket_poller_timeout_ms: int = 100
@@ -75,7 +90,8 @@ class BehaviourSettings(BaseModel):
     # VAD settings
     vad_prob_threshold: float = 0.5
     vad_initial_since_speech: int = 100
-    vad_non_speech_patience_chunks: int = 3
+    vad_non_speech_patience_chunks: int = 15
+    vad_begin_silence_chunks: int = 6
 
     # transcription behaviour
     transcription_max_concurrent_tasks: int = 3
@@ -99,6 +115,8 @@ class LLMSettings(BaseModel):
     :ivar n_parallel: The number of parallel calls allowed to be made to the LLM.
     """
 
+    # ATTENTION: When adding/removing settings, make sure to update the .env.example file
+
     local_llm_url: str = "http://localhost:1234/v1/chat/completions"
     local_llm_model: str = "gpt-oss"
     chat_temperature: float = 1.0
@@ -115,6 +133,8 @@ class VADSettings(BaseModel):
     :ivar sample_rate_hz: Sample rate in Hz for the VAD model.
     """
 
+    # ATTENTION: When adding/removing settings, make sure to update the .env.example file
+
     repo_or_dir: str = "snakers4/silero-vad"
     model_name: str = "silero_vad"
     sample_rate_hz: int = 16000
@@ -128,6 +148,8 @@ class SpeechModelSettings(BaseModel):
     :ivar openai_model_name: Model name for OpenAI-based speech recognition.
     """
 
+    # ATTENTION: When adding/removing settings, make sure to update the .env.example file
+
     # model identifiers for speech recognition
     mlx_model_name: str = "mlx-community/whisper-small.en-mlx"
     openai_model_name: str = "small.en"
@@ -139,6 +161,7 @@ class Settings(BaseSettings):
 
     :ivar app_title: Title of the application.
     :ivar ui_url: URL of the frontend UI.
+    :ivar ri_host: The hostname of the Robot Interface.
     :ivar zmq_settings: ZMQ configuration.
     :ivar agent_settings: Agent name configuration.
     :ivar behaviour_settings: Behavior configuration.
@@ -151,6 +174,8 @@ class Settings(BaseSettings):
 
     ui_url: str = "http://localhost:5173"
 
+    ri_host: str = "localhost"
+
     zmq_settings: ZMQSettings = ZMQSettings()
 
     agent_settings: AgentSettings = AgentSettings()

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/internal_message.py

@@ -12,6 +12,6 @@ class InternalMessage(BaseModel):
     """
 
     to: str
-    sender: str
+    sender: str | None = None
     body: str
     thread: str | None = None

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
 
@@ -180,7 +180,6 @@ class Trigger(ProgramElement):
     :ivar plan: The plan to execute.
     """
 
-    name: str = ""
     condition: Belief
     plan: Plan
 

test/integration/agents/perception/vad_agent/test_vad_agent.py

@@ -91,7 +91,7 @@ def test_out_socket_creation(zmq_context):
     assert per_vad_agent.audio_out_socket is not None
 
     zmq_context.return_value.socket.assert_called_once_with(zmq.PUB)
-    zmq_context.return_value.socket.return_value.bind_to_random_port.assert_called_once()
+    zmq_context.return_value.socket.return_value.bind.assert_called_once_with("inproc://vad_stream")
 
 
 @pytest.mark.asyncio

test/unit/agents/actuation/test_robot_gesture_agent.py

@@ -73,7 +73,7 @@ async def test_setup_connect(zmq_context, mocker):
 async def test_handle_message_sends_valid_gesture_command():
     """Internal message with valid gesture tag is forwarded to robot pub socket."""
     pubsocket = AsyncMock()
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.pubsocket = pubsocket
 
     payload = {
@@ -91,7 +91,7 @@ async def test_handle_message_sends_valid_gesture_command():
 async def test_handle_message_sends_non_gesture_command():
     """Internal message with non-gesture endpoint is not forwarded by this agent."""
     pubsocket = AsyncMock()
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.pubsocket = pubsocket
 
     payload = {"endpoint": "some_other_endpoint", "data": "invalid_tag_not_in_list"}
@@ -107,7 +107,7 @@ async def test_handle_message_sends_non_gesture_command():
 async def test_handle_message_rejects_invalid_gesture_tag():
     """Internal message with invalid gesture tag is not forwarded."""
     pubsocket = AsyncMock()
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.pubsocket = pubsocket
 
     # Use a tag that's not in gesture_data
@@ -123,7 +123,7 @@ async def test_handle_message_rejects_invalid_gesture_tag():
 async def test_handle_message_invalid_payload():
     """Invalid payload is caught and does not send."""
     pubsocket = AsyncMock()
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.pubsocket = pubsocket
 
     msg = InternalMessage(to="robot", sender="tester", body=json.dumps({"bad": "data"}))
@@ -142,12 +142,12 @@ async def test_zmq_command_loop_valid_gesture_payload():
     async def recv_once():
         # stop after first iteration
         agent._running = False
-        return (b"command", json.dumps(command).encode("utf-8"))
+        return b"command", json.dumps(command).encode("utf-8")
 
     fake_socket.recv_multipart = recv_once
     fake_socket.send_json = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.subsocket = fake_socket
     agent.pubsocket = fake_socket
     agent._running = True
@@ -165,12 +165,12 @@ async def test_zmq_command_loop_valid_non_gesture_payload():
 
     async def recv_once():
         agent._running = False
-        return (b"command", json.dumps(command).encode("utf-8"))
+        return b"command", json.dumps(command).encode("utf-8")
 
     fake_socket.recv_multipart = recv_once
     fake_socket.send_json = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.subsocket = fake_socket
     agent.pubsocket = fake_socket
     agent._running = True
@@ -188,12 +188,12 @@ async def test_zmq_command_loop_invalid_gesture_tag():
 
     async def recv_once():
         agent._running = False
-        return (b"command", json.dumps(command).encode("utf-8"))
+        return b"command", json.dumps(command).encode("utf-8")
 
     fake_socket.recv_multipart = recv_once
     fake_socket.send_json = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.subsocket = fake_socket
     agent.pubsocket = fake_socket
     agent._running = True
@@ -210,12 +210,12 @@ async def test_zmq_command_loop_invalid_json():
 
     async def recv_once():
         agent._running = False
-        return (b"command", b"{not_json}")
+        return b"command", b"{not_json}"
 
     fake_socket.recv_multipart = recv_once
     fake_socket.send_json = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.subsocket = fake_socket
     agent.pubsocket = fake_socket
     agent._running = True
@@ -232,12 +232,12 @@ async def test_zmq_command_loop_ignores_send_gestures_topic():
 
     async def recv_once():
         agent._running = False
-        return (b"send_gestures", b"{}")
+        return b"send_gestures", b"{}"
 
     fake_socket.recv_multipart = recv_once
     fake_socket.send_json = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.subsocket = fake_socket
     agent.pubsocket = fake_socket
     agent._running = True
@@ -259,7 +259,9 @@ async def test_fetch_gestures_loop_without_amount():
     fake_repsocket.recv = recv_once
     fake_repsocket.send = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no", "wave", "point"])
+    agent = RobotGestureAgent(
+        "robot_gesture", gesture_data=["hello", "yes", "no", "wave", "point"], address=""
+    )
     agent.repsocket = fake_repsocket
     agent._running = True
 
@@ -287,7 +289,9 @@ async def test_fetch_gestures_loop_with_amount():
     fake_repsocket.recv = recv_once
     fake_repsocket.send = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no", "wave", "point"])
+    agent = RobotGestureAgent(
+        "robot_gesture", gesture_data=["hello", "yes", "no", "wave", "point"], address=""
+    )
     agent.repsocket = fake_repsocket
     agent._running = True
 
@@ -315,7 +319,7 @@ async def test_fetch_gestures_loop_with_integer_request():
     fake_repsocket.recv = recv_once
     fake_repsocket.send = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.repsocket = fake_repsocket
     agent._running = True
 
@@ -340,7 +344,7 @@ async def test_fetch_gestures_loop_with_invalid_json():
     fake_repsocket.recv = recv_once
     fake_repsocket.send = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.repsocket = fake_repsocket
     agent._running = True
 
@@ -365,7 +369,7 @@ async def test_fetch_gestures_loop_with_non_integer_json():
     fake_repsocket.recv = recv_once
     fake_repsocket.send = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.repsocket = fake_repsocket
     agent._running = True
 
@@ -381,7 +385,7 @@ async def test_fetch_gestures_loop_with_non_integer_json():
 def test_gesture_data_attribute():
     """Test that gesture_data returns the expected list."""
     gesture_data = ["hello", "yes", "no", "wave"]
-    agent = RobotGestureAgent("robot_gesture", gesture_data=gesture_data)
+    agent = RobotGestureAgent("robot_gesture", gesture_data=gesture_data, address="")
 
     assert agent.gesture_data == gesture_data
     assert isinstance(agent.gesture_data, list)
@@ -398,7 +402,7 @@ async def test_stop_closes_sockets():
     pubsocket = MagicMock()
     subsocket = MagicMock()
     repsocket = MagicMock()
-    agent = RobotGestureAgent("robot_gesture")
+    agent = RobotGestureAgent("robot_gesture", address="")
     agent.pubsocket = pubsocket
     agent.subsocket = subsocket
     agent.repsocket = repsocket
@@ -415,7 +419,7 @@ async def test_stop_closes_sockets():
 async def test_initialization_with_custom_gesture_data():
     """Agent can be initialized with custom gesture data."""
     custom_gestures = ["custom1", "custom2", "custom3"]
-    agent = RobotGestureAgent("robot_gesture", gesture_data=custom_gestures)
+    agent = RobotGestureAgent("robot_gesture", gesture_data=custom_gestures, address="")
 
     assert agent.gesture_data == custom_gestures
 
@@ -432,7 +436,7 @@ async def test_fetch_gestures_loop_handles_exception():
     fake_repsocket.recv = recv_once
     fake_repsocket.send = AsyncMock()
 
-    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"])
+    agent = RobotGestureAgent("robot_gesture", gesture_data=["hello", "yes", "no"], address="")
     agent.repsocket = fake_repsocket
     agent.logger = MagicMock()
     agent._running = True

test/unit/agents/bdi/test_bdi_program_manager.py

@@ -80,6 +80,7 @@ async def test_receive_programs_valid_and_invalid():
     manager._internal_pub_socket = AsyncMock()
     manager.sub_socket = sub
     manager._create_agentspeak_and_send_to_bdi = AsyncMock()
+    manager._send_clear_llm_history = AsyncMock()
 
     try:
         # Will give StopAsyncIteration when the predefined `sub.recv_multipart` side-effects run out
@@ -92,3 +93,24 @@ async def test_receive_programs_valid_and_invalid():
     forwarded: Program = manager._create_agentspeak_and_send_to_bdi.await_args[0][0]
     assert forwarded.phases[0].norms[0].name == "N1"
     assert forwarded.phases[0].goals[0].name == "G1"
+
+    # Verify history clear was triggered
+    assert manager._send_clear_llm_history.await_count == 1
+
+
+@pytest.mark.asyncio
+async def test_send_clear_llm_history(mock_settings):
+    # Ensure the mock returns a string for the agent name (just like in your LLM tests)
+    mock_settings.agent_settings.llm_agent_name = "llm_agent"
+
+    manager = BDIProgramManager(name="program_manager_test")
+    manager.send = AsyncMock()
+
+    await manager._send_clear_llm_history()
+
+    assert manager.send.await_count == 2
+    msg: InternalMessage = manager.send.await_args_list[0][0][0]
+
+    # Verify the content and recipient
+    assert msg.body == "clear_history"
+    assert msg.to == "llm_agent"

test/unit/agents/bdi/test_text_belief_extractor.py

@@ -6,10 +6,13 @@ import httpx
 import pytest
 
 from control_backend.agents.bdi import TextBeliefExtractorAgent
+from control_backend.agents.bdi.text_belief_extractor_agent import BeliefState
 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_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 (
     ConditionalNorm,
     KeywordBelief,
@@ -23,11 +26,21 @@ from control_backend.schemas.program import (
 
 
 @pytest.fixture
-def agent():
-    agent = TextBeliefExtractorAgent("text_belief_agent")
-    agent.send = AsyncMock()
-    agent._query_llm = AsyncMock()
-    return agent
+def llm():
+    llm = TextBeliefExtractorAgent.LLM(MagicMock(), 4)
+    llm._query_llm = AsyncMock()
+    return llm
+
+
+@pytest.fixture
+def agent(llm):
+    with patch(
+        "control_backend.agents.bdi.text_belief_extractor_agent.TextBeliefExtractorAgent.LLM",
+        return_value=llm,
+    ):
+        agent = TextBeliefExtractorAgent("text_belief_agent")
+        agent.send = AsyncMock()
+        return agent
 
 
 @pytest.fixture
@@ -102,24 +115,12 @@ async def test_handle_message_from_transcriber(agent, mock_settings):
 
     agent.send.assert_awaited_once()  # noqa  # `agent.send` has no such property, but we mock it.
     sent: InternalMessage = agent.send.call_args.args[0]  # noqa
-    assert sent.to == mock_settings.agent_settings.bdi_belief_collector_name
+    assert sent.to == mock_settings.agent_settings.bdi_core_name
     assert sent.thread == "beliefs"
-    parsed = json.loads(sent.body)
-    assert parsed == {"beliefs": {"user_said": [transcription]}, "type": "belief_extraction_text"}
-
-
-@pytest.mark.asyncio
-async def test_process_user_said(agent, mock_settings):
-    transcription = "this is a test"
-
-    await agent._user_said(transcription)
-
-    agent.send.assert_awaited_once()  # noqa  # `agent.send` has no such property, but we mock it.
-    sent: InternalMessage = agent.send.call_args.args[0]  # noqa
-    assert sent.to == mock_settings.agent_settings.bdi_belief_collector_name
-    assert sent.thread == "beliefs"
-    parsed = json.loads(sent.body)
-    assert parsed["beliefs"]["user_said"] == [transcription]
+    parsed = BeliefMessage.model_validate_json(sent.body)
+    replaced_last = parsed.replace.pop()
+    assert replaced_last.name == "user_said"
+    assert replaced_last.arguments == [transcription]
 
 
 @pytest.mark.asyncio
@@ -144,46 +145,46 @@ async def test_query_llm():
         "control_backend.agents.bdi.text_belief_extractor_agent.httpx.AsyncClient",
         return_value=mock_async_client,
     ):
-        agent = TextBeliefExtractorAgent("text_belief_agent")
+        llm = TextBeliefExtractorAgent.LLM(MagicMock(), 4)
 
-        res = await agent._query_llm("hello world", {"type": "null"})
+        res = await llm._query_llm("hello world", {"type": "null"})
         # Response content was set as "null", so should be deserialized as None
         assert res is None
 
 
 @pytest.mark.asyncio
-async def test_retry_query_llm_success(agent):
-    agent._query_llm.return_value = None
-    res = await agent._retry_query_llm("hello world", {"type": "null"})
+async def test_retry_query_llm_success(llm):
+    llm._query_llm.return_value = None
+    res = await llm.query("hello world", {"type": "null"})
 
-    agent._query_llm.assert_called_once()
+    llm._query_llm.assert_called_once()
     assert res is None
 
 
 @pytest.mark.asyncio
-async def test_retry_query_llm_success_after_failure(agent):
-    agent._query_llm.side_effect = [KeyError(), "real value"]
-    res = await agent._retry_query_llm("hello world", {"type": "string"})
+async def test_retry_query_llm_success_after_failure(llm):
+    llm._query_llm.side_effect = [KeyError(), "real value"]
+    res = await llm.query("hello world", {"type": "string"})
 
-    assert agent._query_llm.call_count == 2
+    assert llm._query_llm.call_count == 2
     assert res == "real value"
 
 
 @pytest.mark.asyncio
-async def test_retry_query_llm_failures(agent):
-    agent._query_llm.side_effect = [KeyError(), KeyError(), KeyError(), "real value"]
-    res = await agent._retry_query_llm("hello world", {"type": "string"})
+async def test_retry_query_llm_failures(llm):
+    llm._query_llm.side_effect = [KeyError(), KeyError(), KeyError(), "real value"]
+    res = await llm.query("hello world", {"type": "string"})
 
-    assert agent._query_llm.call_count == 3
+    assert llm._query_llm.call_count == 3
     assert res is None
 
 
 @pytest.mark.asyncio
-async def test_retry_query_llm_fail_immediately(agent):
-    agent._query_llm.side_effect = [KeyError(), "real value"]
-    res = await agent._retry_query_llm("hello world", {"type": "string"}, tries=1)
+async def test_retry_query_llm_fail_immediately(llm):
+    llm._query_llm.side_effect = [KeyError(), "real value"]
+    res = await llm.query("hello world", {"type": "string"}, tries=1)
 
-    assert agent._query_llm.call_count == 1
+    assert llm._query_llm.call_count == 1
     assert res is None
 
 
@@ -192,7 +193,7 @@ async def test_extracting_semantic_beliefs(agent):
     """
     The Program Manager sends beliefs to this agent. Test whether the agent handles them correctly.
     """
-    assert len(agent.available_beliefs) == 0
+    assert len(agent.belief_inferrer.available_beliefs) == 0
     beliefs = BeliefList(
         beliefs=[
             KeywordBelief(
@@ -213,26 +214,28 @@ async def test_extracting_semantic_beliefs(agent):
             to=settings.agent_settings.text_belief_extractor_name,
             sender=settings.agent_settings.bdi_program_manager_name,
             body=beliefs.model_dump_json(),
+            thread="beliefs",
         ),
     )
-    assert len(agent.available_beliefs) == 2
+    assert len(agent.belief_inferrer.available_beliefs) == 2
 
 
 @pytest.mark.asyncio
-async def test_handle_invalid_program(agent, sample_program):
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
-    agent.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
-    assert len(agent.available_beliefs) == 2
+async def test_handle_invalid_beliefs(agent, sample_program):
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
+    assert len(agent.belief_inferrer.available_beliefs) == 2
 
     await agent.handle_message(
         InternalMessage(
             to=settings.agent_settings.text_belief_extractor_name,
             sender=settings.agent_settings.bdi_program_manager_name,
             body=json.dumps({"phases": "Invalid"}),
+            thread="beliefs",
         ),
     )
 
-    assert len(agent.available_beliefs) == 2
+    assert len(agent.belief_inferrer.available_beliefs) == 2
 
 
 @pytest.mark.asyncio
@@ -254,13 +257,13 @@ async def test_handle_robot_response(agent):
 
 
 @pytest.mark.asyncio
-async def test_simulated_real_turn_with_beliefs(agent, sample_program):
+async def test_simulated_real_turn_with_beliefs(agent, llm, sample_program):
     """Test sending user message to extract beliefs from."""
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
-    agent.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
 
     # Send a user message with the belief that there's no more booze
-    agent._query_llm.return_value = {"is_pirate": None, "no_more_booze": True}
+    llm._query_llm.return_value = {"is_pirate": None, "no_more_booze": True}
     assert len(agent.conversation.messages) == 0
     await agent.handle_message(
         InternalMessage(
@@ -275,20 +278,20 @@ async def test_simulated_real_turn_with_beliefs(agent, sample_program):
     assert agent.send.call_count == 2
 
     # First should be the beliefs message
-    message: InternalMessage = agent.send.call_args_list[0].args[0]
+    message: InternalMessage = agent.send.call_args_list[1].args[0]
     beliefs = BeliefMessage.model_validate_json(message.body)
     assert len(beliefs.create) == 1
     assert beliefs.create[0].name == "no_more_booze"
 
 
 @pytest.mark.asyncio
-async def test_simulated_real_turn_no_beliefs(agent, sample_program):
+async def test_simulated_real_turn_no_beliefs(agent, llm, sample_program):
     """Test a user message to extract beliefs from, but no beliefs are formed."""
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
-    agent.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
 
     # Send a user message with no new beliefs
-    agent._query_llm.return_value = {"is_pirate": None, "no_more_booze": None}
+    llm._query_llm.return_value = {"is_pirate": None, "no_more_booze": None}
     await agent.handle_message(
         InternalMessage(
             to=settings.agent_settings.text_belief_extractor_name,
@@ -302,17 +305,17 @@ async def test_simulated_real_turn_no_beliefs(agent, sample_program):
 
 
 @pytest.mark.asyncio
-async def test_simulated_real_turn_no_new_beliefs(agent, sample_program):
+async def test_simulated_real_turn_no_new_beliefs(agent, llm, sample_program):
     """
     Test a user message to extract beliefs from, but no new beliefs are formed because they already
     existed.
     """
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
-    agent.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
-    agent.beliefs["is_pirate"] = True
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
+    agent._current_beliefs = BeliefState(true={InternalBelief(name="is_pirate", arguments=None)})
 
     # Send a user message with the belief the user is a pirate, still
-    agent._query_llm.return_value = {"is_pirate": True, "no_more_booze": None}
+    llm._query_llm.return_value = {"is_pirate": True, "no_more_booze": None}
     await agent.handle_message(
         InternalMessage(
             to=settings.agent_settings.text_belief_extractor_name,
@@ -326,17 +329,19 @@ async def test_simulated_real_turn_no_new_beliefs(agent, sample_program):
 
 
 @pytest.mark.asyncio
-async def test_simulated_real_turn_remove_belief(agent, sample_program):
+async def test_simulated_real_turn_remove_belief(agent, llm, sample_program):
     """
     Test a user message to extract beliefs from, but an existing belief is determined no longer to
     hold.
     """
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
-    agent.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
-    agent.beliefs["no_more_booze"] = True
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
+    agent._current_beliefs = BeliefState(
+        true={InternalBelief(name="no_more_booze", arguments=None)},
+    )
 
     # Send a user message with the belief the user is a pirate, still
-    agent._query_llm.return_value = {"is_pirate": None, "no_more_booze": False}
+    llm._query_llm.return_value = {"is_pirate": None, "no_more_booze": False}
     await agent.handle_message(
         InternalMessage(
             to=settings.agent_settings.text_belief_extractor_name,
@@ -349,18 +354,23 @@ async def test_simulated_real_turn_remove_belief(agent, sample_program):
     assert agent.send.call_count == 2
 
     # Agent's current beliefs should've changed
-    assert not agent.beliefs["no_more_booze"]
+    assert any(b.name == "no_more_booze" for b in agent._current_beliefs.false)
 
 
 @pytest.mark.asyncio
-async def test_llm_failure_handling(agent, sample_program):
+async def test_llm_failure_handling(agent, llm, sample_program):
     """
     Check that the agent handles failures gracefully without crashing.
     """
-    agent._query_llm.side_effect = httpx.HTTPError("")
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
-    agent.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
+    llm._query_llm.side_effect = httpx.HTTPError("")
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.available_beliefs.append(sample_program.phases[0].triggers[0].condition)
 
-    belief_changes = await agent._infer_turn()
+    belief_changes = await agent.belief_inferrer.infer_from_conversation(
+        ChatHistory(
+            messages=[ChatMessage(role="user", content="Good day!")],
+        ),
+    )
 
-    assert len(belief_changes) == 0
+    assert len(belief_changes.true) == 0
+    assert len(belief_changes.false) == 0

test/unit/agents/llm/test_llm_agent.py

@@ -265,3 +265,23 @@ async def test_stream_query_llm_skips_non_data_lines(mock_httpx_client, mock_set
 
     # Only the valid 'data:' line should yield content
     assert tokens == ["Hi"]
+
+
+@pytest.mark.asyncio
+async def test_clear_history_command(mock_settings):
+    """Test that the 'clear_history' message clears the agent's memory."""
+    # setup LLM to have some history
+    mock_settings.agent_settings.bdi_program_manager_name = "bdi_program_manager_agent"
+    agent = LLMAgent("llm_agent")
+    agent.history = [
+        {"role": "user", "content": "Old conversation context"},
+        {"role": "assistant", "content": "Old response"},
+    ]
+    assert len(agent.history) == 2
+    msg = InternalMessage(
+        to="llm_agent",
+        sender=mock_settings.agent_settings.bdi_program_manager_name,
+        body="clear_history",
+    )
+    await agent.handle_message(msg)
+    assert len(agent.history) == 0

test/unit/agents/perception/vad_agent/test_vad_streaming.py

@@ -7,6 +7,15 @@ import zmq
 from control_backend.agents.perception.vad_agent import VADAgent
 
 
+# We don't want to use real ZMQ in unit tests, for example because it can give errors when sockets
+# aren't closed properly.
+@pytest.fixture(autouse=True)
+def mock_zmq():
+    with patch("zmq.asyncio.Context") as mock:
+        mock.instance.return_value = MagicMock()
+        yield mock
+
+
 @pytest.fixture
 def audio_out_socket():
     return AsyncMock()
@@ -140,12 +149,10 @@ async def test_vad_model_load_failure_stops_agent(vad_agent):
         # Patch stop to an AsyncMock so we can check it was awaited
         vad_agent.stop = AsyncMock()
 
-        result = await vad_agent.setup()
+        await vad_agent.setup()
 
         # Assert stop was called
         vad_agent.stop.assert_awaited_once()
-        # Assert setup returned None
-        assert result is None
 
 
 @pytest.mark.asyncio
@@ -155,7 +162,7 @@ async def test_audio_out_bind_failure_sets_none_and_logs(vad_agent, caplog):
     audio_out_socket is set to None, None is returned, and an error is logged.
     """
     mock_socket = MagicMock()
-    mock_socket.bind_to_random_port.side_effect = zmq.ZMQBindError()
+    mock_socket.bind.side_effect = zmq.ZMQBindError()
     with patch("control_backend.agents.perception.vad_agent.azmq.Context.instance") as mock_ctx:
         mock_ctx.return_value.socket.return_value = mock_socket