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_ast.py

@@ -187,9 +187,10 @@ class StatementType(StrEnum):
     EMPTY = ""
     DO_ACTION = "."
     ACHIEVE_GOAL = "!"
-    # TEST_GOAL = "?"  # TODO
+    TEST_GOAL = "?"
     ADD_BELIEF = "+"
     REMOVE_BELIEF = "-"
+    REPLACE_BELIEF = "-+"
 
 
 @dataclass

src/control_backend/agents/bdi/agentspeak_generator.py

@@ -0,0 +1,443 @@
+from functools import singledispatchmethod
+
+from slugify import slugify
+
+from control_backend.agents.bdi.agentspeak_ast import (
+    AstBinaryOp,
+    AstExpression,
+    AstLiteral,
+    AstPlan,
+    AstProgram,
+    AstRule,
+    AstStatement,
+    AstString,
+    AstVar,
+    BinaryOperatorType,
+    StatementType,
+    TriggerType,
+)
+from control_backend.schemas.program import (
+    BasicNorm,
+    ConditionalNorm,
+    GestureAction,
+    Goal,
+    InferredBelief,
+    KeywordBelief,
+    LLMAction,
+    LogicalOperator,
+    Norm,
+    Phase,
+    PlanElement,
+    Program,
+    ProgramElement,
+    SemanticBelief,
+    SpeechAction,
+    Trigger,
+)
+
+
+class AgentSpeakGenerator:
+    _asp: AstProgram
+
+    def generate(self, program: Program) -> str:
+        self._asp = AstProgram()
+
+        self._asp.rules.append(AstRule(self._astify(program.phases[0])))
+        self._add_keyword_inference()
+        self._add_default_plans()
+
+        self._process_phases(program.phases)
+
+        self._add_fallbacks()
+
+        return str(self._asp)
+
+    def _add_keyword_inference(self) -> None:
+        keyword = AstVar("Keyword")
+        message = AstVar("Message")
+        position = AstVar("Pos")
+
+        self._asp.rules.append(
+            AstRule(
+                AstLiteral("keyword_said", [keyword]),
+                AstLiteral("user_said", [message])
+                & AstLiteral(".substring", [keyword, message, position])
+                & (position >= 0),
+            )
+        )
+
+    def _add_default_plans(self):
+        self._add_reply_with_goal_plan()
+        self._add_say_plan()
+        self._add_reply_plan()
+
+    def _add_reply_with_goal_plan(self):
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_GOAL,
+                AstLiteral("reply_with_goal", [AstVar("Goal")]),
+                [AstLiteral("user_said", [AstVar("Message")])],
+                [
+                    AstStatement(StatementType.ADD_BELIEF, AstLiteral("responded_this_turn")),
+                    AstStatement(
+                        StatementType.DO_ACTION,
+                        AstLiteral(
+                            "findall",
+                            [AstVar("Norm"), AstLiteral("norm", [AstVar("Norm")]), AstVar("Norms")],
+                        ),
+                    ),
+                    AstStatement(
+                        StatementType.DO_ACTION,
+                        AstLiteral(
+                            "reply_with_goal", [AstVar("Message"), AstVar("Norms"), AstVar("Goal")]
+                        ),
+                    ),
+                ],
+            )
+        )
+
+    def _add_say_plan(self):
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_GOAL,
+                AstLiteral("say", [AstVar("Text")]),
+                [],
+                [
+                    AstStatement(StatementType.ADD_BELIEF, AstLiteral("responded_this_turn")),
+                    AstStatement(StatementType.DO_ACTION, AstLiteral("say", [AstVar("Text")])),
+                ],
+            )
+        )
+
+    def _add_reply_plan(self):
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_GOAL,
+                AstLiteral("reply"),
+                [AstLiteral("user_said", [AstVar("Message")])],
+                [
+                    AstStatement(StatementType.ADD_BELIEF, AstLiteral("responded_this_turn")),
+                    AstStatement(
+                        StatementType.DO_ACTION,
+                        AstLiteral(
+                            "findall",
+                            [AstVar("Norm"), AstLiteral("norm", [AstVar("Norm")]), AstVar("Norms")],
+                        ),
+                    ),
+                    AstStatement(
+                        StatementType.DO_ACTION,
+                        AstLiteral("reply", [AstVar("Message"), AstVar("Norms")]),
+                    ),
+                ],
+            )
+        )
+
+    def _process_phases(self, phases: list[Phase]) -> None:
+        for curr_phase, next_phase in zip([None] + phases, phases + [None], strict=True):
+            if curr_phase:
+                self._process_phase(curr_phase)
+            self._add_phase_transition(curr_phase, next_phase)
+
+        # End phase behavior
+        # When deleting this, the entire `reply` plan and action can be deleted
+        self._asp.plans.append(
+            AstPlan(
+                type=TriggerType.ADDED_BELIEF,
+                trigger_literal=AstLiteral("user_said", [AstVar("Message")]),
+                context=[AstLiteral("phase", [AstString("end")])],
+                body=[
+                    AstStatement(StatementType.DO_ACTION, AstLiteral("notify_user_said")),
+                    AstStatement(StatementType.ACHIEVE_GOAL, AstLiteral("reply")),
+                ],
+            )
+        )
+
+    def _process_phase(self, phase: Phase) -> None:
+        for norm in phase.norms:
+            self._process_norm(norm, phase)
+
+        self._add_default_loop(phase)
+
+        previous_goal = None
+        for goal in phase.goals:
+            self._process_goal(goal, phase, previous_goal, main_goal=True)
+            previous_goal = goal
+
+        for trigger in phase.triggers:
+            self._process_trigger(trigger, phase)
+
+    def _add_phase_transition(self, from_phase: Phase | None, to_phase: Phase | None) -> None:
+        if from_phase is None:
+            return
+        from_phase_ast = self._astify(from_phase)
+        to_phase_ast = (
+            self._astify(to_phase) if to_phase else AstLiteral("phase", [AstString("end")])
+        )
+
+        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")])
+        #             ),
+        #         ]
+        #     )
+
+        # 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)
+        )
+
+    def _process_norm(self, norm: Norm, phase: Phase) -> None:
+        rule: AstRule | None = None
+
+        match norm:
+            case ConditionalNorm(condition=cond):
+                rule = AstRule(self._astify(norm), self._astify(phase) & self._astify(cond))
+            case BasicNorm():
+                rule = AstRule(self._astify(norm), self._astify(phase))
+
+        if not rule:
+            return
+
+        self._asp.rules.append(rule)
+
+    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")))
+
+        for goal in phase.goals:
+            actions.append(AstStatement(StatementType.ACHIEVE_GOAL, self._astify(goal)))
+
+        actions.append(AstStatement(StatementType.ACHIEVE_GOAL, AstLiteral("transition_phase")))
+
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_BELIEF,
+                AstLiteral("user_said", [AstVar("Message")]),
+                [self._astify(phase)],
+                actions,
+            )
+        )
+
+    def _process_goal(
+        self,
+        goal: Goal,
+        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))
+        if previous_goal and previous_goal.can_fail:
+            context.append(self._astify(previous_goal, achieved=True))
+        if not continues_response:
+            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:
+            body.append(self._step_to_statement(step))
+            if isinstance(step, Goal):
+                subgoals.append(step)
+
+        if not goal.can_fail and not continues_response:
+            body.append(AstStatement(StatementType.ADD_BELIEF, self._astify(goal, achieved=True)))
+
+        self._asp.plans.append(AstPlan(TriggerType.ADDED_GOAL, self._astify(goal), context, body))
+
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_GOAL,
+                self._astify(goal),
+                context=[],
+                body=[AstStatement(StatementType.EMPTY, AstLiteral("true"))],
+            )
+        )
+
+        prev_goal = None
+        for subgoal in subgoals:
+            self._process_goal(subgoal, phase, prev_goal)
+            prev_goal = subgoal
+
+    def _step_to_statement(self, step: PlanElement) -> AstStatement:
+        match step:
+            case Goal() | SpeechAction() | LLMAction() as a:
+                return AstStatement(StatementType.ACHIEVE_GOAL, self._astify(a))
+            case GestureAction() as a:
+                return AstStatement(StatementType.DO_ACTION, self._astify(a))
+
+    # TODO: separate handling of keyword and others
+    def _process_trigger(self, trigger: Trigger, phase: Phase) -> None:
+        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(
+                TriggerType.ADDED_GOAL,
+                AstLiteral("check_triggers"),
+                [self._astify(phase), self._astify(trigger.condition)],
+                body,
+            )
+        )
+
+        # 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)
+
+    def _add_fallbacks(self):
+        # Trigger fallback
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_GOAL,
+                AstLiteral("check_triggers"),
+                [],
+                [AstStatement(StatementType.EMPTY, AstLiteral("true"))],
+            )
+        )
+
+        # Phase transition fallback
+        self._asp.plans.append(
+            AstPlan(
+                TriggerType.ADDED_GOAL,
+                AstLiteral("transition_phase"),
+                [],
+                [AstStatement(StatementType.EMPTY, AstLiteral("true"))],
+            )
+        )
+
+    @singledispatchmethod
+    def _astify(self, element: ProgramElement) -> AstExpression:
+        raise NotImplementedError(f"Cannot convert element {element} to an AgentSpeak expression.")
+
+    @_astify.register
+    def _(self, kwb: KeywordBelief) -> AstExpression:
+        return AstLiteral("keyword_said", [AstString(kwb.keyword)])
+
+    @_astify.register
+    def _(self, sb: SemanticBelief) -> AstExpression:
+        return AstLiteral(self.slugify(sb))
+
+    @_astify.register
+    def _(self, ib: InferredBelief) -> AstExpression:
+        return AstBinaryOp(
+            self._astify(ib.left),
+            BinaryOperatorType.AND if ib.operator == LogicalOperator.AND else BinaryOperatorType.OR,
+            self._astify(ib.right),
+        )
+
+    @_astify.register
+    def _(self, norm: Norm) -> AstExpression:
+        functor = "critical_norm" if norm.critical else "norm"
+        return AstLiteral(functor, [AstString(norm.norm)])
+
+    @_astify.register
+    def _(self, phase: Phase) -> AstExpression:
+        return AstLiteral("phase", [AstString(str(phase.id))])
+
+    @_astify.register
+    def _(self, goal: Goal, achieved: bool = False) -> AstExpression:
+        return AstLiteral(f"{'achieved_' if achieved else ''}{self._slugify_str(goal.name)}")
+
+    @_astify.register
+    def _(self, trigger: Trigger) -> AstExpression:
+        return AstLiteral(self.slugify(trigger))
+
+    @_astify.register
+    def _(self, sa: SpeechAction) -> AstExpression:
+        return AstLiteral("say", [AstString(sa.text)])
+
+    @_astify.register
+    def _(self, ga: GestureAction) -> AstExpression:
+        gesture = ga.gesture
+        return AstLiteral("gesture", [AstString(gesture.type), AstString(gesture.name)])
+
+    @_astify.register
+    def _(self, la: LLMAction) -> AstExpression:
+        return AstLiteral("reply_with_goal", [AstString(la.goal)])
+
+    @singledispatchmethod
+    @staticmethod
+    def slugify(element: ProgramElement) -> str:
+        raise NotImplementedError(f"Cannot convert element {element} to a slug.")
+
+    @slugify.register
+    @staticmethod
+    def _(sb: SemanticBelief) -> str:
+        return f"semantic_{AgentSpeakGenerator._slugify_str(sb.name)}"
+
+    @slugify.register
+    @staticmethod
+    def _(g: Goal) -> str:
+        return AgentSpeakGenerator._slugify_str(g.name)
+
+    @slugify.register
+    @staticmethod
+    def _(t: Trigger):
+        return f"trigger_{AgentSpeakGenerator._slugify_str(t.name)}"
+
+    @staticmethod
+    def _slugify_str(text: str) -> str:
+        return slugify(text, separator="_", stopwords=["a", "an", "the", "we", "you", "I"])

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
 
@@ -42,13 +43,13 @@ class BDICoreAgent(BaseAgent):
 
     bdi_agent: agentspeak.runtime.Agent
 
-    def __init__(self, name: str, asl: str):
+    def __init__(self, name: str):
         super().__init__(name)
-        self.asl_file = asl
         self.env = agentspeak.runtime.Environment()
         # Deep copy because we don't actually want to modify the standard actions globally
         self.actions = copy.deepcopy(agentspeak.stdlib.actions)
         self._wake_bdi_loop = asyncio.Event()
+        self._bdi_loop_task = None
 
     async def setup(self) -> None:
         """
@@ -65,19 +66,22 @@ class BDICoreAgent(BaseAgent):
         await self._load_asl()
 
         # Start the BDI cycle loop
-        self.add_behavior(self._bdi_loop())
+        self._bdi_loop_task = self.add_behavior(self._bdi_loop())
         self._wake_bdi_loop.set()
         self.logger.debug("Setup complete.")
 
-    async def _load_asl(self):
+    async def _load_asl(self, file_name: str | None = None) -> None:
         """
         Load and parse the AgentSpeak source file.
         """
+        file_name = file_name or "src/control_backend/agents/bdi/default_behavior.asl"
+
         try:
-            with open(self.asl_file) as source:
+            with open(file_name) as source:
                 self.bdi_agent = self.env.build_agent(source, self.actions)
+                self.logger.info(f"Loaded new ASL from {file_name}.")
         except FileNotFoundError:
-            self.logger.warning(f"Could not find the specified ASL file at {self.asl_file}.")
+            self.logger.warning(f"Could not find the specified ASL file at {file_name}.")
             self.bdi_agent = agentspeak.runtime.Agent(self.env, self.name)
 
     async def _bdi_loop(self):
@@ -97,7 +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
 
@@ -116,6 +119,7 @@ class BDICoreAgent(BaseAgent):
         Handle incoming messages.
 
         - **Beliefs**: Updates the internal belief base.
+        - **Program**: Updates the internal agentspeak file to match the current program.
         - **LLM Responses**: Forwards the generated text to the Robot Speech Agent (actuation).
 
         :param msg: The received internal message.
@@ -130,6 +134,13 @@ class BDICoreAgent(BaseAgent):
                 self.logger.exception("Error processing belief.")
             return
 
+        # New agentspeak file
+        if msg.thread == "new_program":
+            if self._bdi_loop_task:
+                self._bdi_loop_task.cancel()
+            await self._load_asl(msg.body)
+            self.add_behavior(self._bdi_loop())
+
         # The message was not a belief, handle special cases based on sender
         match msg.sender:
             case settings.agent_settings.llm_name:
@@ -144,6 +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):
         """
@@ -190,14 +212,33 @@ 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.
         """
+        if args is None:
+            term = agentspeak.Literal(name)
+        else:
             new_args = (agentspeak.Literal(arg) for arg in args)
             term = agentspeak.Literal(name, new_args)
 
@@ -239,6 +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
@@ -246,20 +318,18 @@ class BDICoreAgent(BaseAgent):
         the function expects (which will be located in `term.args`).
         """
 
-        @self.actions.add(".reply", 3)
+        @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.
             """
             message_text = agentspeak.grounded(term.args[0], intention.scope)
             norms = agentspeak.grounded(term.args[1], intention.scope)
-            goals = agentspeak.grounded(term.args[2], intention.scope)
 
             self.logger.debug("Norms: %s", norms)
-            self.logger.debug("Goals: %s", goals)
             self.logger.debug("User text: %s", message_text)
 
-            asyncio.create_task(self._send_to_llm(str(message_text), str(norms), str(goals)))
+            self.add_behavior(self._send_to_llm(str(message_text), str(norms), ""))
             yield
 
         @self.actions.add(".reply_with_goal", 3)
@@ -278,11 +348,11 @@ class BDICoreAgent(BaseAgent):
                 norms,
                 goal,
             )
-            # asyncio.create_task(self._send_to_llm(str(message_text), norms, str(goal)))
+            self.add_behavior(self._send_to_llm(str(message_text), str(norms), str(goal)))
             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.
             """
@@ -290,17 +360,27 @@ class BDICoreAgent(BaseAgent):
 
             self.logger.debug('"say" action called with text=%s', message_text)
 
-            # speech_command = SpeechCommand(data=message_text)
+            speech_command = SpeechCommand(data=message_text)
-            # speech_message = InternalMessage(
+            speech_message = InternalMessage(
-            #     to=settings.agent_settings.robot_speech_name,
+                to=settings.agent_settings.robot_speech_name,
-            #     sender=settings.agent_settings.bdi_core_name,
+                sender=settings.agent_settings.bdi_core_name,
-            #     body=speech_command.model_dump_json(),
+                body=speech_command.model_dump_json(),
-            # )
+            )
-            # asyncio.create_task(agent.send(speech_message))
+
+            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.
             """
@@ -313,13 +393,113 @@ class BDICoreAgent(BaseAgent):
                 gesture_name,
             )
 
-            # gesture = Gesture(type=gesture_type, name=gesture_name)
+            if str(gesture_type) == "single":
-            # gesture_message = InternalMessage(
+                endpoint = RIEndpoint.GESTURE_SINGLE
-            #     to=settings.agent_settings.robot_gesture_name,
+            elif str(gesture_type) == "tag":
-            #     sender=settings.agent_settings.bdi_core_name,
+                endpoint = RIEndpoint.GESTURE_TAG
-            #     body=gesture.model_dump_json(),
+            else:
-            # )
+                self.logger.warning("Gesture type %s could not be resolved.", gesture_type)
-            # asyncio.create_task(agent.send(gesture_message))
+                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):
@@ -331,13 +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

@@ -1,599 +1,24 @@
-import uuid
+import asyncio
-from collections.abc import Iterable
+import json
 
 import zmq
 from pydantic import ValidationError
-from slugify import slugify
 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, GoalList
+from control_backend.schemas.internal_message import InternalMessage
 from control_backend.schemas.program import (
-    Action,
-    BasicBelief,
-    BasicNorm,
     Belief,
     ConditionalNorm,
-    GestureAction,
     Goal,
     InferredBelief,
-    KeywordBelief,
-    LLMAction,
-    LogicalOperator,
     Phase,
-    Plan,
     Program,
-    ProgramElement,
-    SemanticBelief,
-    SpeechAction,
-    Trigger,
 )
 
-test_program = Program(
-    phases=[
-        Phase(
-            norms=[
-                BasicNorm(norm="Talk like a pirate", id=uuid.uuid4()),
-                ConditionalNorm(
-                    condition=InferredBelief(
-                        left=KeywordBelief(keyword="Arr", id=uuid.uuid4()),
-                        right=SemanticBelief(
-                            description="testing", name="semantic belief", id=uuid.uuid4()
-                        ),
-                        operator=LogicalOperator.OR,
-                        name="Talking to a pirate",
-                        id=uuid.uuid4(),
-                    ),
-                    norm="Use nautical terms",
-                    id=uuid.uuid4(),
-                ),
-                ConditionalNorm(
-                    condition=SemanticBelief(
-                        description="We are talking to a child",
-                        name="talking to child",
-                        id=uuid.uuid4(),
-                    ),
-                    norm="Do not use cuss words",
-                    id=uuid.uuid4(),
-                ),
-            ],
-            triggers=[
-                Trigger(
-                    condition=InferredBelief(
-                        left=KeywordBelief(keyword="key", id=uuid.uuid4()),
-                        right=InferredBelief(
-                            left=KeywordBelief(keyword="key2", id=uuid.uuid4()),
-                            right=SemanticBelief(
-                                description="Decode this", name="semantic belief 2", id=uuid.uuid4()
-                            ),
-                            operator=LogicalOperator.OR,
-                            name="test trigger inferred inner",
-                            id=uuid.uuid4(),
-                        ),
-                        operator=LogicalOperator.OR,
-                        name="test trigger inferred outer",
-                        id=uuid.uuid4(),
-                    ),
-                    plan=Plan(
-                        steps=[
-                            SpeechAction(text="Testing trigger", id=uuid.uuid4()),
-                            Goal(
-                                name="Testing trigger",
-                                plan=Plan(
-                                    steps=[LLMAction(goal="Do something", id=uuid.uuid4())],
-                                    id=uuid.uuid4(),
-                                ),
-                                id=uuid.uuid4(),
-                            ),
-                        ],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                )
-            ],
-            goals=[
-                Goal(
-                    name="Determine user age",
-                    plan=Plan(
-                        steps=[LLMAction(goal="Determine the age of the user.", id=uuid.uuid4())],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                ),
-                Goal(
-                    name="Find the user's name",
-                    plan=Plan(
-                        steps=[
-                            Goal(
-                                name="Greet the user",
-                                plan=Plan(
-                                    steps=[LLMAction(goal="Greet the user.", id=uuid.uuid4())],
-                                    id=uuid.uuid4(),
-                                ),
-                                can_fail=False,
-                                id=uuid.uuid4(),
-                            ),
-                            Goal(
-                                name="Ask for name",
-                                plan=Plan(
-                                    steps=[
-                                        LLMAction(goal="Obtain the user's name.", id=uuid.uuid4())
-                                    ],
-                                    id=uuid.uuid4(),
-                                ),
-                                id=uuid.uuid4(),
-                            ),
-                        ],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                ),
-                Goal(
-                    name="Tell a joke",
-                    plan=Plan(
-                        steps=[LLMAction(goal="Tell a joke.", id=uuid.uuid4())], id=uuid.uuid4()
-                    ),
-                    id=uuid.uuid4(),
-                ),
-            ],
-            id=uuid.uuid4(),
-        ),
-        Phase(
-            id=uuid.uuid4(),
-            norms=[
-                BasicNorm(norm="Use very gentle speech.", id=uuid.uuid4()),
-                ConditionalNorm(
-                    condition=SemanticBelief(
-                        description="We are talking to a child",
-                        name="talking to child",
-                        id=uuid.uuid4(),
-                    ),
-                    norm="Do not use cuss words",
-                    id=uuid.uuid4(),
-                ),
-            ],
-            triggers=[
-                Trigger(
-                    condition=InferredBelief(
-                        left=KeywordBelief(keyword="help", id=uuid.uuid4()),
-                        right=SemanticBelief(
-                            description="User is stuck", name="stuck", id=uuid.uuid4()
-                        ),
-                        operator=LogicalOperator.OR,
-                        name="help_or_stuck",
-                        id=uuid.uuid4(),
-                    ),
-                    plan=Plan(
-                        steps=[
-                            Goal(
-                                name="Unblock user",
-                                plan=Plan(
-                                    steps=[
-                                        LLMAction(
-                                            goal="Provide a step-by-step path to "
-                                            "resolve the user's issue.",
-                                            id=uuid.uuid4(),
-                                        )
-                                    ],
-                                    id=uuid.uuid4(),
-                                ),
-                                id=uuid.uuid4(),
-                            ),
-                        ],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                ),
-            ],
-            goals=[
-                Goal(
-                    name="Clarify intent",
-                    plan=Plan(
-                        steps=[
-                            LLMAction(
-                                goal="Ask 1-2 targeted questions to clarify the "
-                                "user's intent, then proceed.",
-                                id=uuid.uuid4(),
-                            )
-                        ],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                ),
-                Goal(
-                    name="Provide solution",
-                    plan=Plan(
-                        steps=[
-                            LLMAction(
-                                goal="Deliver a solution to complete the user's goal.",
-                                id=uuid.uuid4(),
-                            )
-                        ],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                ),
-                Goal(
-                    name="Summarize next steps",
-                    plan=Plan(
-                        steps=[
-                            LLMAction(
-                                goal="Summarize what the user should do next.", id=uuid.uuid4()
-                            )
-                        ],
-                        id=uuid.uuid4(),
-                    ),
-                    id=uuid.uuid4(),
-                ),
-            ],
-        ),
-    ]
-)
-
-
-def do_things():
-    print(AgentSpeakGenerator().generate(test_program))
-
-
-class AgentSpeakGenerator:
-    """
-    Converts Pydantic representation of behavior programs into AgentSpeak(L) code string.
-    """
-
-    arrow_prefix = f"{' ' * 2}<-{' ' * 2}"
-    colon_prefix = f"{' ' * 2}:{' ' * 3}"
-    indent_prefix = " " * 6
-
-    def generate(self, program: Program) -> str:
-        lines = []
-        lines.append("")
-
-        lines += self._generate_initial_beliefs(program)
-
-        lines += self._generate_basic_flow(program)
-
-        lines += self._generate_phase_transitions(program)
-
-        lines += self._generate_norms(program)
-
-        lines += self._generate_belief_inference(program)
-
-        lines += self._generate_goals(program)
-
-        lines += self._generate_triggers(program)
-
-        return "\n".join(lines)
-
-    def _generate_initial_beliefs(self, program: Program) -> Iterable[str]:
-        yield "// --- Initial beliefs and agent startup ---"
-
-        yield "phase(start)."
-
-        yield ""
-
-        yield "+started"
-        yield f"{self.colon_prefix}phase(start)"
-        yield f"{self.arrow_prefix}phase({program.phases[0].id if program.phases else 'end'})."
-
-        yield from ["", ""]
-
-    def _generate_basic_flow(self, program: Program) -> Iterable[str]:
-        yield "// --- Basic flow ---"
-
-        for phase in program.phases:
-            yield from self._generate_basic_flow_per_phase(phase)
-
-        yield from ["", ""]
-
-    def _generate_basic_flow_per_phase(self, phase: Phase) -> Iterable[str]:
-        yield "+user_said(Message)"
-        yield f"{self.colon_prefix}phase({phase.id})"
-
-        goals = phase.goals
-        if goals:
-            yield f"{self.arrow_prefix}{self._slugify(goals[0], include_prefix=True)}"
-            for goal in goals[1:]:
-                yield f"{self.indent_prefix}{self._slugify(goal, include_prefix=True)}"
-
-        yield f"{self.indent_prefix if goals else self.arrow_prefix}!transition_phase."
-
-    def _generate_phase_transitions(self, program: Program) -> Iterable[str]:
-        yield "// --- Phase transitions ---"
-
-        if len(program.phases) == 0:
-            yield from ["", ""]
-            return
-
-        # TODO: remove outdated things
-
-        for i in range(-1, len(program.phases)):
-            predecessor = program.phases[i] if i >= 0 else None
-            successor = program.phases[i + 1] if i < len(program.phases) - 1 else None
-            yield from self._generate_phase_transition(predecessor, successor)
-
-        yield from self._generate_phase_transition(None, None)  # to avoid failing plan
-
-        yield from ["", ""]
-
-    def _generate_phase_transition(
-        self, phase: Phase | None = None, next_phase: Phase | None = None
-    ) -> Iterable[str]:
-        yield "+!transition_phase"
-
-        if phase is None and next_phase is None:  # base case true to avoid failing plan
-            yield f"{self.arrow_prefix}true."
-            return
-
-        yield f"{self.colon_prefix}phase({phase.id if phase else 'start'})"
-        yield f"{self.arrow_prefix}-+phase({next_phase.id if next_phase else 'end'})."
-
-    def _generate_norms(self, program: Program) -> Iterable[str]:
-        yield "// --- Norms ---"
-
-        for phase in program.phases:
-            for norm in phase.norms:
-                if type(norm) is BasicNorm:
-                    yield f"{self._slugify(norm)} :- phase({phase.id})."
-                if type(norm) is ConditionalNorm:
-                    yield (
-                        f"{self._slugify(norm)} :- phase({phase.id}) & "
-                        f"{self._slugify(norm.condition)}."
-                    )
-
-        yield from ["", ""]
-
-    def _generate_belief_inference(self, program: Program) -> Iterable[str]:
-        yield "// --- Belief inference rules ---"
-
-        for phase in program.phases:
-            for norm in phase.norms:
-                if not isinstance(norm, ConditionalNorm):
-                    continue
-
-                yield from self._belief_inference_recursive(norm.condition)
-
-            for trigger in phase.triggers:
-                yield from self._belief_inference_recursive(trigger.condition)
-
-        yield from ["", ""]
-
-    def _belief_inference_recursive(self, belief: Belief) -> Iterable[str]:
-        if type(belief) is KeywordBelief:
-            yield (
-                f"{self._slugify(belief)} :- user_said(Message) & "
-                f'.substring(Message, "{belief.keyword}", Pos) & Pos >= 0.'
-            )
-        if type(belief) is InferredBelief:
-            yield (
-                f"{self._slugify(belief)} :- {self._slugify(belief.left)} "
-                f"{'&' if belief.operator == LogicalOperator.AND else '|'} "
-                f"{self._slugify(belief.right)}."
-            )
-
-            yield from self._belief_inference_recursive(belief.left)
-            yield from self._belief_inference_recursive(belief.right)
-
-    def _generate_goals(self, program: Program) -> Iterable[str]:
-        yield "// --- Goals ---"
-
-        for phase in program.phases:
-            previous_goal: Goal | None = None
-            for goal in phase.goals:
-                yield from self._generate_goal_plan_recursive(goal, phase, previous_goal)
-                previous_goal = goal
-
-        yield from ["", ""]
-
-    def _generate_goal_plan_recursive(
-        self, goal: Goal, phase: Phase, previous_goal: Goal | None = None
-    ) -> Iterable[str]:
-        yield f"+{self._slugify(goal, include_prefix=True)}"
-
-        # Context
-        yield f"{self.colon_prefix}phase({phase.id}) &"
-        yield f"{self.indent_prefix}not responded_this_turn &"
-        yield f"{self.indent_prefix}not achieved_{self._slugify(goal)} &"
-        if previous_goal:
-            yield f"{self.indent_prefix}achieved_{self._slugify(previous_goal)}"
-        else:
-            yield f"{self.indent_prefix}true"
-
-        extra_goals_to_generate = []
-
-        steps = goal.plan.steps
-
-        if len(steps) == 0:
-            yield f"{self.arrow_prefix}true."
-            return
-
-        first_step = steps[0]
-        yield (
-            f"{self.arrow_prefix}{self._slugify(first_step, include_prefix=True)}"
-            f"{'.' if len(steps) == 1 and goal.can_fail else ';'}"
-        )
-        if isinstance(first_step, Goal):
-            extra_goals_to_generate.append(first_step)
-
-        for step in steps[1:-1]:
-            yield f"{self.indent_prefix}{self._slugify(step, include_prefix=True)};"
-            if isinstance(step, Goal):
-                extra_goals_to_generate.append(step)
-
-        if len(steps) > 1:
-            last_step = steps[-1]
-            yield (
-                f"{self.indent_prefix}{self._slugify(last_step, include_prefix=True)}"
-                f"{'.' if goal.can_fail else ';'}"
-            )
-            if isinstance(last_step, Goal):
-                extra_goals_to_generate.append(last_step)
-
-        if not goal.can_fail:
-            yield f"{self.indent_prefix}+achieved_{self._slugify(goal)}."
-
-        yield f"+{self._slugify(goal, include_prefix=True)}"
-        yield f"{self.arrow_prefix}true."
-
-        yield ""
-
-        extra_previous_goal: Goal | None = None
-        for extra_goal in extra_goals_to_generate:
-            yield from self._generate_goal_plan_recursive(extra_goal, phase, extra_previous_goal)
-            extra_previous_goal = extra_goal
-
-    def _generate_triggers(self, program: Program) -> Iterable[str]:
-        yield "// --- Triggers ---"
-
-        for phase in program.phases:
-            for trigger in phase.triggers:
-                yield from self._generate_trigger_plan(trigger, phase)
-
-        yield from ["", ""]
-
-    def _generate_trigger_plan(self, trigger: Trigger, phase: Phase) -> Iterable[str]:
-        belief_name = self._slugify(trigger.condition)
-
-        yield f"+{belief_name}"
-        yield f"{self.colon_prefix}phase({phase.id})"
-
-        extra_goals_to_generate = []
-
-        steps = trigger.plan.steps
-
-        if len(steps) == 0:
-            yield f"{self.arrow_prefix}true."
-            return
-
-        first_step = steps[0]
-        yield (
-            f"{self.arrow_prefix}{self._slugify(first_step, include_prefix=True)}"
-            f"{'.' if len(steps) == 1 else ';'}"
-        )
-        if isinstance(first_step, Goal):
-            extra_goals_to_generate.append(first_step)
-
-        for step in steps[1:-1]:
-            yield f"{self.indent_prefix}{self._slugify(step, include_prefix=True)};"
-            if isinstance(step, Goal):
-                extra_goals_to_generate.append(step)
-
-        if len(steps) > 1:
-            last_step = steps[-1]
-            yield f"{self.indent_prefix}{self._slugify(last_step, include_prefix=True)}."
-            if isinstance(last_step, Goal):
-                extra_goals_to_generate.append(last_step)
-
-        yield ""
-
-        extra_previous_goal: Goal | None = None
-        for extra_goal in extra_goals_to_generate:
-            yield from self._generate_trigger_plan_recursive(extra_goal, phase, extra_previous_goal)
-            extra_previous_goal = extra_goal
-
-    def _generate_trigger_plan_recursive(
-        self, goal: Goal, phase: Phase, previous_goal: Goal | None = None
-    ) -> Iterable[str]:
-        yield f"+{self._slugify(goal, include_prefix=True)}"
-
-        extra_goals_to_generate = []
-
-        steps = goal.plan.steps
-
-        if len(steps) == 0:
-            yield f"{self.arrow_prefix}true."
-            return
-
-        first_step = steps[0]
-        yield (
-            f"{self.arrow_prefix}{self._slugify(first_step, include_prefix=True)}"
-            f"{'.' if len(steps) == 1 and goal.can_fail else ';'}"
-        )
-        if isinstance(first_step, Goal):
-            extra_goals_to_generate.append(first_step)
-
-        for step in steps[1:-1]:
-            yield f"{self.indent_prefix}{self._slugify(step, include_prefix=True)};"
-            if isinstance(step, Goal):
-                extra_goals_to_generate.append(step)
-
-        if len(steps) > 1:
-            last_step = steps[-1]
-            yield (
-                f"{self.indent_prefix}{self._slugify(last_step, include_prefix=True)}"
-                f"{'.' if goal.can_fail else ';'}"
-            )
-            if isinstance(last_step, Goal):
-                extra_goals_to_generate.append(last_step)
-
-        if not goal.can_fail:
-            yield f"{self.indent_prefix}+achieved_{self._slugify(goal)}."
-
-        yield f"+{self._slugify(goal, include_prefix=True)}"
-        yield f"{self.arrow_prefix}true."
-
-        yield ""
-
-        extra_previous_goal: Goal | None = None
-        for extra_goal in extra_goals_to_generate:
-            yield from self._generate_goal_plan_recursive(extra_goal, phase, extra_previous_goal)
-            extra_previous_goal = extra_goal
-
-    def _slugify(self, element: ProgramElement, include_prefix: bool = False) -> str:
-        def base_slugify_call(text: str):
-            return slugify(text, separator="_", stopwords=["a", "the"])
-
-        if type(element) is KeywordBelief:
-            return f'keyword_said("{element.keyword}")'
-
-        if type(element) is SemanticBelief:
-            name = element.name
-            return f"semantic_{base_slugify_call(name if name else element.description)}"
-
-        if isinstance(element, BasicNorm):
-            return f'norm("{element.norm}")'
-
-        if isinstance(element, Goal):
-            return f"{'!' if include_prefix else ''}{base_slugify_call(element.name)}"
-
-        if isinstance(element, SpeechAction):
-            return f'.say("{element.text}")'
-
-        if isinstance(element, GestureAction):
-            return f'.gesture("{element.gesture}")'
-
-        if isinstance(element, LLMAction):
-            return f'!generate_response_with_goal("{element.goal}")'
-
-        if isinstance(element, Action.__value__):
-            raise NotImplementedError(
-                "Have not implemented an ASL string representation for this action."
-            )
-
-        if element.name == "":
-            raise ValueError("Name must be initialized for this type of ProgramElement.")
-
-        return base_slugify_call(element.name)
-
-    def _extract_basic_beliefs_from_program(self, program: Program) -> list[BasicBelief]:
-        beliefs = []
-
-        for phase in program.phases:
-            for norm in phase.norms:
-                if isinstance(norm, ConditionalNorm):
-                    beliefs += self._extract_basic_beliefs_from_belief(norm.condition)
-
-            for trigger in phase.triggers:
-                beliefs += self._extract_basic_beliefs_from_belief(trigger.condition)
-
-        return beliefs
-
-    def _extract_basic_beliefs_from_belief(self, belief: Belief) -> list[BasicBelief]:
-        if isinstance(belief, InferredBelief):
-            return self._extract_basic_beliefs_from_belief(
-                belief.left
-            ) + self._extract_basic_beliefs_from_belief(belief.right)
-        return [belief]
-
 
 class BDIProgramManager(BaseAgent):
     """
@@ -607,44 +32,161 @@ 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
 
-    # async def _send_to_bdi(self, program: Program):
+    def _initialize_internal_state(self, program: Program):
-    #     """
+        self._program = program
-    #     Convert a received program into BDI beliefs and send them to the BDI Core Agent.
+        self._phase = program.phases[0]  # start in first phase
-    #
+
-    #     Currently, it takes the **first phase** of the program and extracts:
+    async def _create_agentspeak_and_send_to_bdi(self, program: Program):
-    #     -   **Norms**: Constraints or rules the agent must follow.
+        """
-    #     -   **Goals**: Objectives the agent must achieve.
+        Convert a received program into an AgentSpeak file and send it to the BDI Core Agent.
-    #
+
-    #     These are sent as a ``BeliefMessage`` with ``replace=True``, meaning they will
+        :param program: The program object received from the API.
-    #     overwrite any existing norms/goals of the same name in the BDI agent.
+        """
-    #
+        asg = AgentSpeakGenerator()
-    #     :param program: The program object received from the API.
+
-    #     """
+        asl_str = asg.generate(program)
-    #     first_phase = program.phases[0]
+
-    #     norms_belief = Belief(
+        file_name = "src/control_backend/agents/bdi/agentspeak.asl"
-    #         name="norms",
+
-    #         arguments=[norm.norm for norm in first_phase.norms],
+        with open(file_name, "w") as f:
-    #         replace=True,
+            f.write(asl_str)
-    #     )
+
-    #     goals_belief = Belief(
+        msg = InternalMessage(
-    #         name="goals",
+            sender=self.name,
-    #         arguments=[goal.description for goal in first_phase.goals],
+            to=settings.agent_settings.bdi_core_name,
-    #         replace=True,
+            body=file_name,
-    #     )
+            thread="new_program",
-    #     program_beliefs = BeliefMessage(beliefs=[norms_belief, goals_belief])
+        )
-    #
+
-    #     message = InternalMessage(
+        await self.send(msg)
-    #         to=settings.agent_settings.bdi_core_name,
+
-    #         sender=self.name,
+    async def handle_message(self, msg: InternalMessage):
-    #         body=program_beliefs.model_dump_json(),
+        match msg.thread:
-    #         thread="beliefs",
+            case "transition_phase":
-    #     )
+                phases = json.loads(msg.body)
-    #     await self.send(message)
+
-    #     self.logger.debug("Sent new norms and goals to the BDI agent.")
+                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 norm in self._phase.norms:
+            if isinstance(norm, ConditionalNorm):
+                beliefs += self._extract_beliefs_from_belief(norm.condition)
+
+        for trigger in self._phase.triggers:
+            beliefs += self._extract_beliefs_from_belief(trigger.condition)
+
+        return beliefs
+
+    @staticmethod
+    def _extract_beliefs_from_belief(belief: Belief) -> list[Belief]:
+        if isinstance(belief, InferredBelief):
+            return BDIProgramManager._extract_beliefs_from_belief(
+                belief.left
+            ) + BDIProgramManager._extract_beliefs_from_belief(belief.right)
+        return [belief]
+
+    async def _send_beliefs_to_semantic_belief_extractor(self):
+        """
+        Extract beliefs from the program and send them to the Semantic Belief Extractor Agent.
+        """
+        beliefs = BeliefList(beliefs=self._extract_current_beliefs())
+
+        message = InternalMessage(
+            to=settings.agent_settings.text_belief_extractor_name,
+            sender=self.name,
+            body=beliefs.model_dump_json(),
+            thread="beliefs",
+        )
+
+        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):
         """
@@ -652,6 +194,7 @@ class BDIProgramManager(BaseAgent):
 
         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()
@@ -659,10 +202,18 @@ 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
 
-            await self._send_to_bdi(program)
+            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(),
+                self._send_goals_to_semantic_belief_extractor(),
+            )
 
     async def setup(self):
         """
@@ -678,7 +229,3 @@ class BDIProgramManager(BaseAgent):
         self.sub_socket.subscribe("program")
 
         self.add_behavior(self._receive_programs())
-
-
-if __name__ == "__main__":
-    do_things()

src/control_backend/agents/bdi/default_behavior.asl

@@ -0,0 +1,6 @@
+norms("").
+
++user_said(Message) : norms(Norms) <-
+  .notify_user_said(Message);
+  -user_said(Message);
+  .reply(Message, Norms).

src/control_backend/agents/bdi/rules.asl

@@ -1,6 +0,0 @@
-norms("").
-goals("").
-
-+user_said(Message) : norms(Norms) & goals(Goals) <-
-  -user_said(Message);
-  .reply(Message, Norms, Goals).

src/control_backend/agents/bdi/text_belief_extractor_agent.py

@@ -2,23 +2,46 @@ import asyncio
 import json
 
 import httpx
-from pydantic import ValidationError
+from pydantic import BaseModel, ValidationError
-from slugify import slugify
 
 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, 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 (
+from control_backend.schemas.program import Goal, SemanticBelief
-    Belief,
+
-    ConditionalNorm,
+type JSONLike = None | bool | int | float | str | list["JSONLike"] | dict[str, "JSONLike"]
-    InferredBelief,
+
-    Program,
+
-    SemanticBelief,
+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):
     """
@@ -34,8 +57,11 @@ class TextBeliefExtractorAgent(BaseAgent):
 
     def __init__(self, name: str):
         super().__init__(name)
-        self.beliefs: dict[str, bool] = {}
+        self._llm = self.LLM(self, settings.llm_settings.n_parallel)
-        self.available_beliefs: list[SemanticBelief] = []
+        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=[])
 
     async def setup(self):
@@ -57,13 +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
@@ -78,51 +105,66 @@ 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:
-            program = Program.model_validate_json(msg.body)
+            belief_list = BeliefList.model_validate_json(msg.body)
         except ValidationError:
             self.logger.warning(
-                "Received message from program manager but it is not a valid program."
+                "Received message from program manager but it is not a valid list of beliefs."
             )
             return
 
-        self.logger.debug("Received a program from the program manager.")
+        available_beliefs = [b for b in belief_list.beliefs if isinstance(b, SemanticBelief)]
-
+        self.belief_inferrer.available_beliefs = available_beliefs
-        self.available_beliefs = self._extract_basic_beliefs_from_program(program)
+        self.logger.debug(
-
+            "Received %d semantic beliefs from the program manager: %s",
-    # TODO Copied from an incomplete version of the program manager. Use that one instead.
+            len(available_beliefs),
-    @staticmethod
+            ", ".join(b.name for b in available_beliefs),
-    def _extract_basic_beliefs_from_program(program: Program) -> list[SemanticBelief]:
-        beliefs = []
-
-        for phase in program.phases:
-            for norm in phase.norms:
-                if isinstance(norm, ConditionalNorm):
-                    beliefs += TextBeliefExtractorAgent._extract_basic_beliefs_from_belief(
-                        norm.condition
         )
 
-            for trigger in phase.triggers:
+    def _handle_goals_message(self, msg: InternalMessage):
-                beliefs += TextBeliefExtractorAgent._extract_basic_beliefs_from_belief(
+        try:
-                    trigger.condition
+            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
 
-        return beliefs
+        # 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]
-    # TODO Copied from an incomplete version of the program manager. Use that one instead.
+        self.goal_inferrer.goals = available_goals
-    @staticmethod
+        self.logger.debug(
-    def _extract_basic_beliefs_from_belief(belief: Belief) -> list[SemanticBelief]:
+            "Received %d failable goals from the program manager: %s",
-        if isinstance(belief, InferredBelief):
+            len(available_goals),
-            return TextBeliefExtractorAgent._extract_basic_beliefs_from_belief(
+            ", ".join(g.name for g in available_goals),
-                belief.left
+        )
-            ) + TextBeliefExtractorAgent._extract_basic_beliefs_from_belief(belief.right)
-        return [belief]
 
     async def _user_said(self, text: str):
         """
@@ -130,161 +172,88 @@ 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):
-        """
+        conversation_beliefs = await self.belief_inferrer.infer_from_conversation(self.conversation)
-        Process conversation history to extract beliefs, semantically. Any changed beliefs are sent
+
-        to the BDI core.
+        new_beliefs = conversation_beliefs - self._current_beliefs
-        """
+        if not new_beliefs:
-        # Return instantly if there are no beliefs to infer
+            self.logger.debug("No new beliefs detected.")
-        if not self.available_beliefs:
             return
 
-        candidate_beliefs = await self._infer_turn()
+        self._current_beliefs |= new_beliefs
-        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.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)
+        message = InternalMessage(
-            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(
             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
+    async def _infer_goal_completions(self):
-    def _split_into_chunks[T](items: list[T], n: int) -> list[list[T]]:
+        goal_completions = await self.goal_inferrer.infer_from_conversation(self.conversation)
-        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_turn(self) -> dict:
+        new_achieved = [
-        """
+            InternalBelief(name=goal, arguments=None)
-        Process the stored conversation history to extract semantic beliefs. Returns a list of
+            for goal, achieved in goal_completions.items()
-        beliefs that have been set to ``True``, ``False`` or ``None``.
+            if achieved and self._current_goal_completions.get(goal) != achieved
-
-        :return: A dict mapping belief names to a value ``True``, ``False`` or ``None``.
-        """
-        n_parallel = max(1, min(settings.llm_settings.n_parallel - 1, len(self.available_beliefs)))
-        all_beliefs = await asyncio.gather(
-            *[
-                self._infer_beliefs(self.conversation, beliefs)
-                for beliefs in self._split_into_chunks(self.available_beliefs, n_parallel)
         ]
-        )
+        new_not_achieved = [
-        retval = {}
+            InternalBelief(name=goal, arguments=None)
-        for beliefs in all_beliefs:
+            for goal, achieved in goal_completions.items()
-            if beliefs is None:
+            if not achieved and self._current_goal_completions.get(goal) != achieved
-                continue
-            retval.update(beliefs)
-        return retval
-
-    @staticmethod
-    def _create_belief_schema(belief: SemanticBelief) -> tuple[str, dict]:
-        # TODO: use real belief names
-        return belief.name or slugify(belief.description), {
-            "type": ["boolean", "null"],
-            "description": belief.description,
-        }
-
-    @staticmethod
-    def _create_beliefs_schema(beliefs: list[SemanticBelief]) -> dict:
-        belief_schemas = [
-            TextBeliefExtractorAgent._create_belief_schema(belief) for belief in beliefs
         ]
+        for goal, achieved in goal_completions.items():
+            self._current_goal_completions[goal] = achieved
 
-        return {
+        if not new_achieved and not new_not_achieved:
-            "type": "object",
+            self.logger.debug("No goal achievement changes detected.")
-            "properties": dict(belief_schemas),
+            return
-            "required": [name for name, _ in belief_schemas],
-        }
 
-    @staticmethod
+        belief_changes = BeliefMessage(
-    def _format_message(message: ChatMessage):
+            create=new_achieved,
-        return f"{message.role.upper()}:\n{message.content}"
+            delete=new_not_achieved,
-
-    @staticmethod
-    def _format_conversation(conversation: ChatHistory):
-        return "\n\n".join(
-            [TextBeliefExtractorAgent._format_message(message) for message in conversation.messages]
         )
-
+        message = InternalMessage(
-    @staticmethod
+            to=settings.agent_settings.bdi_core_name,
-    def _format_beliefs(beliefs: list[SemanticBelief]):
+            sender=self.name,
-        # TODO: use real belief names
+            body=belief_changes.model_dump_json(),
-        return "\n".join(
+            thread="beliefs",
-            [
-                f"- {belief.name or slugify(belief.description)}: {belief.description}"
-                for belief in beliefs
-            ]
         )
+        await self.send(message)
 
-    async def _infer_beliefs(
+    class LLM:
-        self,
-        conversation: ChatHistory,
-        beliefs: list[SemanticBelief],
-    ) -> dict | None:
         """
-        Infer given beliefs based on the given conversation.
+        Class that handles sending structured generation requests to an LLM.
-        :param conversation: The conversation to infer beliefs from.
-        :param beliefs: The beliefs to infer.
-        :return: A dict containing belief names and a boolean whether they hold, or None if the
-            belief cannot be inferred based on the given conversation.
-        """
-        example = {
-            "example_belief": True,
-        }
-
-        prompt = f"""{self._format_conversation(conversation)}
-
-Given the above conversation, what beliefs can be inferred?
-If there is no relevant information about a belief belief, give null.
-In case messages conflict, prefer using the most recent messages for inference.
-
-Choose from the following list of beliefs, formatted as (belief_name, description):
-{self._format_beliefs(beliefs)}
-
-Respond with a JSON similar to the following, but with the property names as given above:
-{json.dumps(example, indent=2)}
         """
 
-        schema = self._create_beliefs_schema(beliefs)
+        def __init__(self, agent: "TextBeliefExtractorAgent", n_parallel: int):
+            self._agent = agent
+            self._semaphore = asyncio.Semaphore(n_parallel)
 
-        return await self._retry_query_llm(prompt, schema)
+        async def query(self, prompt: str, schema: dict, tries: int = 3) -> JSONLike | None:
-
-    async def _retry_query_llm(self, prompt: str, schema: dict, tries: int = 3) -> dict | 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
@@ -296,7 +265,7 @@ Respond with a JSON similar to the following, but with the property names as giv
                 except (httpx.HTTPError, json.JSONDecodeError, KeyError) as e:
                     if try_count < tries:
                         continue
-                self.logger.exception(
+                    self._agent.logger.exception(
                         "Failed to get LLM response after %d tries.",
                         try_count,
                         exc_info=e,
@@ -304,11 +273,10 @@ Respond with a JSON similar to the following, but with the property names as giv
 
             return None
 
-    @staticmethod
+        async def _query_llm(self, prompt: str, schema: dict) -> JSONLike:
-    async def _query_llm(prompt: str, schema: dict) -> dict:
             """
-        Query an LLM with the given prompt and schema, return an instance of a dict conforming to
+            Query an LLM with the given prompt and schema, return an instance of a dict conforming
-        that schema.
+            to that schema.
 
             :param prompt: The prompt to be queried.
             :param schema: Schema to use during response.
@@ -316,8 +284,10 @@ Respond with a JSON similar to the following, but with the property names as giv
             :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.
+            :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,
@@ -336,10 +306,177 @@ Respond with a JSON similar to the following, but with the property names as giv
                             "temperature": settings.llm_settings.code_temperature,
                             "stream": False,
                         },
-                timeout=None,
+                        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: list[dict[str, bool | None] | None] = await asyncio.gather(
+            *[
+                self._infer_beliefs(conversation, beliefs)
+                for beliefs in self._split_into_chunks(self.available_beliefs, n_parallel)
+            ]
+        )
+        retval = BeliefState()
+        for beliefs in all_beliefs:
+            if beliefs is None:
+                continue
+            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 _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.
+
+        :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[str, bool | None] | None:
+        """
+        Infer given beliefs based on the given conversation.
+        :param conversation: The conversation to infer beliefs from.
+        :param beliefs: The beliefs to infer.
+        :return: A dict containing belief names and a boolean whether they hold, or None if the
+            belief cannot be inferred based on the given conversation.
+        """
+        example = {
+            "example_belief": True,
+        }
+
+        prompt = f"""{self._format_conversation(conversation)}
+
+Given the above conversation, what beliefs can be inferred?
+If there is no relevant information about a belief belief, give null.
+In case messages conflict, prefer using the most recent messages for inference.
+
+Choose from the following list of beliefs, formatted as `- <belief_name>: <description>`:
+{self._format_beliefs(beliefs)}
+
+Respond with a JSON similar to the following, but with the property names as given above:
+{json.dumps(example, indent=2)}
+"""
+
+        schema = self._create_beliefs_schema(beliefs)
+
+        return await self._llm.query(prompt, schema)
+
+    @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]:
+        """
+        Determine which goals have been achieved based on the given conversation.
+
+        :param conversation: The conversation to infer goal completion from.
+        :return: A mapping of goals and a boolean whether they have been achieved.
+        """
+        if not self.goals:
+            return {}
+
+        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)
+        }
+
+    async def _infer_goal(self, conversation: ChatHistory, goal: Goal) -> bool:
+        prompt = f"""{self._format_conversation(conversation)}
+
+Given the above conversation, what has the following goal been achieved?
+
+The name of the goal: {goal.name}
+Description of the goal: {goal.description}
+
+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

@@ -3,14 +3,11 @@ import json
 
 import zmq
 import zmq.asyncio as azmq
-from pydantic import ValidationError
 from zmq.asyncio import Context
 
 from control_backend.agents import BaseAgent
 from control_backend.agents.actuation.robot_gesture_agent import RobotGestureAgent
-from control_backend.core.agent_system import InternalMessage
 from control_backend.core.config import settings
-from control_backend.schemas.ri_message import PauseCommand
 
 from ..actuation.robot_speech_agent import RobotSpeechAgent
 from ..perception import VADAgent
@@ -41,7 +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)
@@ -171,7 +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}"
 
@@ -251,6 +248,7 @@ class RICommunicationAgent(BaseAgent):
                     self._req_socket.recv_json(), timeout=seconds_to_wait_total / 2
                 )
 
+                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.")
@@ -301,11 +299,3 @@ class RICommunicationAgent(BaseAgent):
         self.logger.debug("Restarting communication negotiation.")
         if await self._negotiate_connection(max_retries=1):
             self.connected = True
-
-    async def handle_message(self, msg : InternalMessage):
-        try:
-            pause_command = PauseCommand.model_validate_json(msg.body)
-            self._req_socket.send_json(pause_command.model_dump())
-            self.logger.debug(self._req_socket.recv_json())
-        except ValidationError:
-            self.logger.warning("Incorrect message format for PauseCommand.")

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.")
+            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/mock_agents/test_pause_ri.py

@@ -1,68 +0,0 @@
-import asyncio
-import json
-
-import zmq
-from zmq.asyncio import Context
-
-from control_backend.agents.base import BaseAgent
-from control_backend.core.agent_system import InternalMessage
-from control_backend.core.config import settings
-
-
-class TestPauseAgent(BaseAgent):
-    def __init__(self, name: str):
-        super().__init__(name)
-        
-    async def setup(self):
-        context = Context.instance()
-        self.pub_socket = context.socket(zmq.PUB)
-        self.pub_socket.connect(settings.zmq_settings.internal_pub_address)
-
-        self.add_behavior(self._pause_command_loop())
-        self.logger.debug("TestPauseAgent setup complete.")
-    
-    async def _pause_command_loop(self):
-        print("Starting Pause command test loop.")
-        while True:
-            pause_command = {
-                "endpoint": "pause",
-                "data": True,
-            }
-
-            message = InternalMessage(
-                to="ri_communication_agent",
-                sender=self.name,
-                body=json.dumps(pause_command),
-            )
-            await self.send(message)
-
-            # User interrupt message 
-            data = {
-                "type": "pause",
-                "context": True,
-            }
-            await self.pub_socket.send_multipart([b"button_pressed", json.dumps(data).encode()])
-
-            self.logger.info("Pausing robot actions.")
-            await asyncio.sleep(15)  # Simulate delay between messages
-
-            pause_command = {
-                "endpoint": "pause",
-                "data": False,
-            }
-            message = InternalMessage(
-                to="ri_communication_agent",
-                sender=self.name,
-                body=json.dumps(pause_command),
-            )
-            await self.send(message)
-
-            # User interrupt message
-            data = {
-                "type": "pause",
-                "context": False,
-            }
-            await self.pub_socket.send_multipart([b"button_pressed", json.dumps(data).encode()])
-
-            self.logger.info("Resuming robot actions.")
-            await asyncio.sleep(15)  # Simulate delay between messages

src/control_backend/agents/perception/vad_agent.py

@@ -7,7 +7,6 @@ import zmq.asyncio as azmq
 
 from control_backend.agents import BaseAgent
 from control_backend.core.config import settings
-from control_backend.schemas.internal_message import InternalMessage
 
 from ...schemas.program_status import PROGRAM_STATUS, ProgramStatus
 from .transcription_agent.transcription_agent import TranscriptionAgent
@@ -87,12 +86,6 @@ class VADAgent(BaseAgent):
         self.audio_buffer = np.array([], dtype=np.float32)
         self.i_since_speech = settings.behaviour_settings.vad_initial_since_speech
         self._ready = asyncio.Event()
-
-        # Pause control
-        self._reset_needed = False
-        self._paused = asyncio.Event()
-        self._paused.set()  # Not paused at start
-
         self.model = None
 
     async def setup(self):
@@ -110,12 +103,11 @@ class VADAgent(BaseAgent):
 
         self._connect_audio_in_socket()
 
-        audio_out_port = self._connect_audio_out_socket()
+        audio_out_address = self._connect_audio_out_socket()
-        if audio_out_port is None:
+        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)
@@ -168,13 +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
@@ -220,16 +213,6 @@ class VADAgent(BaseAgent):
         """
         await self._ready.wait()
         while self._running:
-            await self._paused.wait()
-
-            # After being unpaused, reset stream and buffers
-            if self._reset_needed:
-                self.logger.debug("Resuming: resetting stream and buffers.")
-                await self._reset_stream()
-                self.audio_buffer = np.array([], dtype=np.float32)
-                self.i_since_speech = settings.behaviour_settings.vad_initial_since_speech
-                self._reset_needed = False
-
             assert self.audio_in_poller is not None
             data = await self.audio_in_poller.poll()
             if data is None:
@@ -246,10 +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
@@ -263,35 +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.
+            # At this point, we know that there is no speech.
-            # Prepend the last chunk that had no speech, for a more fluent boundary
+            # Prepend the last few chunks that had no speech, for a more fluent boundary.
-            self.audio_buffer = chunk
+            self.audio_buffer = np.append(self.audio_buffer, chunk)
-
+            self.audio_buffer = self.audio_buffer[-begin_silence_length * len(chunk) :]
-    async def handle_message(self, msg: InternalMessage):
-        """
-        Handle incoming messages.
-
-        Expects messages to pause or resume the VAD processing from User Interrupt Agent.
-
-        :param msg: The received internal message.
-        """
-        sender = msg.sender
-
-        if sender == settings.agent_settings.user_interrupt_name:
-            if msg.body == "PAUSE":
-                self.logger.info("Pausing VAD processing.")
-                self._paused.clear()
-                # If the robot needs to pick up speaking where it left off, do not set _reset_needed
-                self._reset_needed = True
-            elif msg.body == "RESUME":
-                self.logger.info("Resuming VAD processing.")
-                self._paused.set()
-            else:
-                self.logger.warning(f"Unknown command from User Interrupt Agent: {msg.body}")
-        else:
-            self.logger.debug(f"Ignoring message from unknown sender: {sender}")

src/control_backend/agents/user_interrupt/user_interrupt_agent.py

@@ -6,12 +6,7 @@ from zmq.asyncio import Context
 from control_backend.agents import BaseAgent
 from control_backend.core.agent_system import InternalMessage
 from control_backend.core.config import settings
-from control_backend.schemas.ri_message import (
+from control_backend.schemas.ri_message import GestureCommand, RIEndpoint, SpeechCommand
-    GestureCommand,
-    PauseCommand,
-    RIEndpoint,
-    SpeechCommand,
-)
 
 
 class UserInterruptAgent(BaseAgent):
@@ -76,12 +71,6 @@ class UserInterruptAgent(BaseAgent):
                     "Forwarded button press (override) with context '%s' to BDIProgramManager.",
                     event_context,
                 )
-            elif event_type == "pause":
-                await self._send_pause_command(event_context)
-                if event_context:
-                    self.logger.info("Sent pause command.")
-                else:
-                    self.logger.info("Sent resume command.")
             else:
                 self.logger.warning(
                     "Received button press with unknown type '%s' (context: '%s').",
@@ -141,38 +130,6 @@ class UserInterruptAgent(BaseAgent):
             belief_id,
         )
 
-    async def _send_pause_command(self, pause : bool):
-        """
-        Send a pause command to the Robot Interface via the RI Communication Agent.
-        Send a pause command to the other internal agents; for now just VAD agent. 
-        """
-        cmd = PauseCommand(data=pause)
-        message = InternalMessage(
-            to=settings.agent_settings.ri_communication_name,
-            sender=self.name,
-            body=cmd.model_dump_json(),
-        )
-        await self.send(message)
-
-        if pause:
-            # Send pause to VAD agent
-            vad_message = InternalMessage(
-                to=settings.agent_settings.vad_name,
-                sender=self.name,
-                body="PAUSE",
-            )
-            await self.send(vad_message)
-            self.logger.info("Sent pause command to VAD Agent and RI Communication Agent.")
-        else:
-            # Send resume to VAD agent
-            vad_message = InternalMessage(
-                to=settings.agent_settings.vad_name,
-                sender=self.name,
-                body="RESUME",
-            )
-            await self.send(vad_message)
-            self.logger.info("Sent resume command to VAD Agent and RI Communication Agent.")
-
     async def setup(self):
         """
         Initialize the agent.

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 the endpoint that the Robot Interface connects to.
-    :ivar ri_communication_address: Address for receiving communication from the Robot Interface.
+    :ivar vad_pub_address: Address that the VAD agent binds to and publishes audio segments to.
-    :ivar vad_agent_address: Address for the Voice Activity Detection (VAD) agent.
     """
 
+    # 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/main.py

@@ -39,11 +39,10 @@ from control_backend.agents.communication import RICommunicationAgent
 # LLM Agents
 from control_backend.agents.llm import LLMAgent
 
-# Other backend imports
-from control_backend.agents.mock_agents.test_pause_ri import TestPauseAgent
-
 # User Interrupt Agent
 from control_backend.agents.user_interrupt.user_interrupt_agent import UserInterruptAgent
+
+# Other backend imports
 from control_backend.api.v1.router import api_router
 from control_backend.core.config import settings
 from control_backend.logging import setup_logging
@@ -121,7 +120,6 @@ async def lifespan(app: FastAPI):
             BDICoreAgent,
             {
                 "name": settings.agent_settings.bdi_core_name,
-                "asl": "src/control_backend/agents/bdi/rules.asl",
             },
         ),
         "BeliefCollectorAgent": (
@@ -142,12 +140,6 @@ async def lifespan(app: FastAPI):
                 "name": settings.agent_settings.bdi_program_manager_name,
             },
         ),
-        "TestPauseAgent": (
-            TestPauseAgent,
-            {
-                "name": "pause_test_agent",
-            },
-        ),
         "UserInterruptAgent": (
             UserInterruptAgent,
             {

src/control_backend/schemas/belief_list.py

@@ -0,0 +1,19 @@
+from pydantic import BaseModel
+
+from control_backend.schemas.program import Belief as ProgramBelief
+from control_backend.schemas.program import Goal
+
+
+class BeliefList(BaseModel):
+    """
+    Represents a list of beliefs, separated from a program. Useful in agents which need to
+    communicate beliefs.
+
+    :ivar: beliefs: The list of beliefs.
+    """
+
+    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

@@ -43,7 +43,6 @@ class SemanticBelief(ProgramElement):
     :ivar description: Description of how to form the belief, used by the LLM.
     """
 
-    name: str = ""
     description: str
 
 
@@ -113,10 +112,12 @@ class Goal(ProgramElement):
     for example when the achieving of the goal is dependent on the user's reply, this means
     that the achieved status will be set from somewhere else in the program.
 
+    :ivar description: A description of the goal, used to determine if it has been achieved.
     :ivar plan: The plan to execute.
     :ivar can_fail: Whether we can fail to achieve the goal after executing the plan.
     """
 
+    description: str = ""
     plan: Plan
     can_fail: bool = True
 
@@ -179,7 +180,6 @@ class Trigger(ProgramElement):
     :ivar plan: The plan to execute.
     """
 
-    name: str = ""
     condition: Belief
     plan: Plan
 

src/control_backend/schemas/ri_message.py

@@ -14,7 +14,6 @@ class RIEndpoint(str, Enum):
     GESTURE_TAG = "actuate/gesture/tag"
     PING = "ping"
     NEGOTIATE_PORTS = "negotiate/ports"
-    PAUSE = "pause"
 
 
 class RIMessage(BaseModel):
@@ -65,14 +64,3 @@ class GestureCommand(RIMessage):
         if self.endpoint not in allowed:
             raise ValueError("endpoint must be GESTURE_SINGLE or GESTURE_TAG")
         return self
-
-class PauseCommand(RIMessage):
-    """
-    A specific command to pause or unpause the robot's actions.
-
-    :ivar endpoint: Fixed to ``RIEndpoint.PAUSE``.
-    :ivar data: A boolean indicating whether to pause (True) or unpause (False).
-    """
-
-    endpoint: RIEndpoint = RIEndpoint(RIEndpoint.PAUSE)
-    data: bool

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_core_agent.py

@@ -20,7 +20,7 @@ def mock_agentspeak_env():
 
 @pytest.fixture
 def agent():
-    agent = BDICoreAgent("bdi_agent", "dummy.asl")
+    agent = BDICoreAgent("bdi_agent")
     agent.send = AsyncMock()
     agent.bdi_agent = MagicMock()
     return agent
@@ -133,14 +133,14 @@ async def test_custom_actions(agent):
 
     # Invoke action
     mock_term = MagicMock()
-    mock_term.args = ["Hello", "Norm", "Goal"]
+    mock_term.args = ["Hello", "Norm"]
     mock_intention = MagicMock()
 
     # Run generator
     gen = action_fn(agent, mock_term, mock_intention)
     next(gen)  # Execute
 
-    agent._send_to_llm.assert_called_with("Hello", "Norm", "Goal")
+    agent._send_to_llm.assert_called_with("Hello", "Norm", "")
 
 
 def test_add_belief_sets_event(agent):

test/unit/agents/bdi/test_bdi_program_manager.py

@@ -32,6 +32,8 @@ def make_valid_program_json(norm="N1", goal="G1") -> str:
                     Goal(
                         id=uuid.uuid4(),
                         name=goal,
+                        description="This description can be used to determine whether the goal "
+                        "has been achieved.",
                         plan=Plan(
                             id=uuid.uuid4(),
                             name="Goal Plan",
@@ -53,7 +55,7 @@ async def test_send_to_bdi():
     manager.send = AsyncMock()
 
     program = Program.model_validate_json(make_valid_program_json())
-    await manager._send_to_bdi(program)
+    await manager._create_agentspeak_and_send_to_bdi(program)
 
     assert manager.send.await_count == 1
     msg: InternalMessage = manager.send.await_args[0][0]
@@ -75,8 +77,10 @@ async def test_receive_programs_valid_and_invalid():
     ]
 
     manager = BDIProgramManager(name="program_manager_test")
+    manager._internal_pub_socket = AsyncMock()
     manager.sub_socket = sub
-    manager._send_to_bdi = AsyncMock()
+    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
@@ -85,7 +89,28 @@ async def test_receive_programs_valid_and_invalid():
         pass
 
     # Only valid Program should have triggered _send_to_bdi
-    assert manager._send_to_bdi.await_count == 1
+    assert manager._create_agentspeak_and_send_to_bdi.await_count == 1
-    forwarded: Program = manager._send_to_bdi.await_args[0][0]
+    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

@@ -6,11 +6,16 @@ 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,
     LLMAction,
     Phase,
     Plan,
@@ -21,10 +26,20 @@ from control_backend.schemas.program import (
 
 
 @pytest.fixture
-def 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()
-    agent._query_llm = AsyncMock()
         return agent
 
 
@@ -100,24 +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)
+    parsed = BeliefMessage.model_validate_json(sent.body)
-    assert parsed == {"beliefs": {"user_said": [transcription]}, "type": "belief_extraction_text"}
+    replaced_last = parsed.replace.pop()
-
+    assert replaced_last.name == "user_said"
-
+    assert replaced_last.arguments == [transcription]
-@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]
 
 
 @pytest.mark.asyncio
@@ -142,77 +145,97 @@ 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):
+async def test_retry_query_llm_success(llm):
-    agent._query_llm.return_value = None
+    llm._query_llm.return_value = None
-    res = await agent._retry_query_llm("hello world", {"type": "null"})
+    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):
+async def test_retry_query_llm_success_after_failure(llm):
-    agent._query_llm.side_effect = [KeyError(), "real value"]
+    llm._query_llm.side_effect = [KeyError(), "real value"]
-    res = await agent._retry_query_llm("hello world", {"type": "string"})
+    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):
+async def test_retry_query_llm_failures(llm):
-    agent._query_llm.side_effect = [KeyError(), KeyError(), KeyError(), "real value"]
+    llm._query_llm.side_effect = [KeyError(), KeyError(), KeyError(), "real value"]
-    res = await agent._retry_query_llm("hello world", {"type": "string"})
+    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):
+async def test_retry_query_llm_fail_immediately(llm):
-    agent._query_llm.side_effect = [KeyError(), "real value"]
+    llm._query_llm.side_effect = [KeyError(), "real value"]
-    res = await agent._retry_query_llm("hello world", {"type": "string"}, tries=1)
+    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
 
 
 @pytest.mark.asyncio
-async def test_extracting_beliefs_from_program(agent, sample_program):
+async def test_extracting_semantic_beliefs(agent):
-    assert len(agent.available_beliefs) == 0
+    """
+    The Program Manager sends beliefs to this agent. Test whether the agent handles them correctly.
+    """
+    assert len(agent.belief_inferrer.available_beliefs) == 0
+    beliefs = BeliefList(
+        beliefs=[
+            KeywordBelief(
+                id=uuid.uuid4(),
+                name="keyword_hello",
+                keyword="hello",
+            ),
+            SemanticBelief(
+                id=uuid.uuid4(), name="semantic_hello_1", description="Some semantic belief 1"
+            ),
+            SemanticBelief(
+                id=uuid.uuid4(), name="semantic_hello_2", description="Some semantic belief 2"
+            ),
+        ]
+    )
     await agent.handle_message(
         InternalMessage(
             to=settings.agent_settings.text_belief_extractor_name,
             sender=settings.agent_settings.bdi_program_manager_name,
-            body=sample_program.model_dump_json(),
+            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):
+async def test_handle_invalid_beliefs(agent, sample_program):
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.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].triggers[0].condition)
-    assert len(agent.available_beliefs) == 2
+    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
@@ -234,13 +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.belief_inferrer.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].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(
@@ -255,20 +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.belief_inferrer.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].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,
@@ -282,17 +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.belief_inferrer.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].triggers[0].condition)
-    agent.beliefs["is_pirate"] = True
+    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,
@@ -306,17 +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.belief_inferrer.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].triggers[0].condition)
-    agent.beliefs["no_more_booze"] = True
+    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,
@@ -329,18 +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("")
+    llm._query_llm.side_effect = httpx.HTTPError("")
-    agent.available_beliefs.append(sample_program.phases[0].norms[0].condition)
+    agent.belief_inferrer.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].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
 

test/unit/api/v1/endpoints/test_program_endpoint.py

@@ -43,6 +43,8 @@ def make_valid_program_dict():
                     Goal(
                         id=uuid.uuid4(),
                         name="Some goal",
+                        description="This description can be used to determine whether the goal "
+                        "has been achieved.",
                         plan=Plan(
                             id=uuid.uuid4(),
                             name="Goal Plan",

test/unit/schemas/test_ui_program_message.py

@@ -31,6 +31,7 @@ def base_goal() -> Goal:
     return Goal(
         id=uuid.uuid4(),
         name="testGoalName",
+        description="This description can be used to determine whether the goal has been achieved.",
         plan=Plan(
             id=uuid.uuid4(),
             name="testGoalPlanName",