Merge branch 'feat/longer-pauses-possible' into 'dev'

Stop LLM response when user adds something to their message

See merge request ics/sp/2025/n25b/pepperplus-cb!47

src/control_backend/agents/bdi/bdi_core_agent.py

@@ -338,7 +338,7 @@ class BDICoreAgent(BaseAgent):
             yield
 
         @self.actions.add(".reply_with_goal", 3)
-        def _reply_with_goal(agent: "BDICoreAgent", term, intention):
+        def _reply_with_goal(agent, term, intention):
             """
             Let the LLM generate a response to a user's utterance with the current norms and a
             specific goal.
@@ -512,10 +512,6 @@ class BDICoreAgent(BaseAgent):
 
             yield
 
-        @self.actions.add(".notify_ui", 0)
-        def _notify_ui(agent, term, intention):
-            pass
-
     async def _send_to_llm(self, text: str, norms: str, goals: str):
         """
         Sends a text query to the LLM agent asynchronously.

src/control_backend/agents/bdi/text_belief_extractor_agent.py

@@ -345,6 +345,9 @@ class TextBeliefExtractorAgent(BaseAgent):
                 async with httpx.AsyncClient() as client:
                     response = await client.post(
                         settings.llm_settings.local_llm_url,
+                        headers={"Authorization": f"Bearer {settings.llm_settings.api_key}"}
+                        if settings.llm_settings.api_key
+                        else {},
                         json={
                             "model": settings.llm_settings.local_llm_model,
                             "messages": [{"role": "user", "content": prompt}],

src/control_backend/agents/llm/llm_agent.py

@@ -1,3 +1,4 @@
+import asyncio
 import json
 import re
 import uuid
@@ -32,6 +33,10 @@ class LLMAgent(BaseAgent):
     def __init__(self, name: str):
         super().__init__(name)
         self.history = []
+        self._querying = False
+        self._interrupted = False
+        self._interrupted_message = ""
+        self._go_ahead = asyncio.Event()
 
     async def setup(self):
         self.logger.info("Setting up %s.", self.name)
@@ -50,13 +55,13 @@ class LLMAgent(BaseAgent):
                 case "prompt_message":
                     try:
                         prompt_message = LLMPromptMessage.model_validate_json(msg.body)
-                        await self._process_bdi_message(prompt_message)
+                        self.add_behavior(self._process_bdi_message(prompt_message))  # no block
                     except ValidationError:
                         self.logger.debug("Prompt message from BDI core is invalid.")
                 case "assistant_message":
-                    self.history.append({"role": "assistant", "content": msg.body})
+                    self._apply_conversation_message({"role": "assistant", "content": msg.body})
                 case "user_message":
-                    self.history.append({"role": "user", "content": msg.body})
+                    self._apply_conversation_message({"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.")
@@ -73,12 +78,45 @@ class LLMAgent(BaseAgent):
 
         :param message: The parsed prompt message containing text, norms, and goals.
         """
+        if self._querying:
+            self.logger.debug("Received another BDI prompt while processing previous message.")
+            self._interrupted = True  # interrupt the previous processing
+            await self._go_ahead.wait()  # wait until we get the go-ahead
+
+        message.text = f"{self._interrupted_message} {message.text}"
+
+        self._go_ahead.clear()
+        self._querying = True
         full_message = ""
         async for chunk in self._query_llm(message.text, message.norms, message.goals):
+            if self._interrupted:
+                self._interrupted_message = message.text
+                self.logger.debug("Interrupted processing of previous message.")
+                break
             await self._send_reply(chunk)
             full_message += chunk
-        self.logger.debug("Finished processing BDI message. Response sent in chunks to BDI core.")
-        await self._send_full_reply(full_message)
+        else:
+            self._querying = False
+
+            self._apply_conversation_message(
+                {
+                    "role": "assistant",
+                    "content": full_message,
+                }
+            )
+            self.logger.debug(
+                "Finished processing BDI message. Response sent in chunks to BDI core."
+            )
+            await self._send_full_reply(full_message)
+
+        self._go_ahead.set()
+        self._interrupted = False
+
+    def _apply_conversation_message(self, message: dict[str, str]):
+        if len(self.history) > 0 and message["role"] == self.history[-1]["role"]:
+            self.history[-1]["content"] += " " + message["content"]
+            return
+        self.history.append(message)
 
     async def _send_reply(self, msg: str):
         """
@@ -159,13 +197,6 @@ class LLMAgent(BaseAgent):
             # Yield any remaining tail
             if current_chunk:
                 yield current_chunk
-
-            self.history.append(
-                {
-                    "role": "assistant",
-                    "content": full_message,
-                }
-            )
         except httpx.HTTPError as err:
             self.logger.error("HTTP error.", exc_info=err)
             yield "LLM service unavailable."
@@ -185,6 +216,9 @@ class LLMAgent(BaseAgent):
             async with client.stream(
                 "POST",
                 settings.llm_settings.local_llm_url,
+                headers={"Authorization": f"Bearer {settings.llm_settings.api_key}"}
+                if settings.llm_settings.api_key
+                else {},
                 json={
                     "model": settings.llm_settings.local_llm_model,
                     "messages": messages,

src/control_backend/agents/perception/transcription_agent/speech_recognizer.py

@@ -145,4 +145,6 @@ class OpenAIWhisperSpeechRecognizer(SpeechRecognizer):
 
     def recognize_speech(self, audio: np.ndarray) -> str:
         self.load_model()
-        return whisper.transcribe(self.model, audio, **self._get_decode_options(audio))["text"]
+        return whisper.transcribe(self.model, audio, **self._get_decode_options(audio))[
+            "text"
+        ].strip()

src/control_backend/core/config.py

@@ -123,6 +123,7 @@ class LLMSettings(BaseModel):
 
     local_llm_url: str = "http://localhost:1234/v1/chat/completions"
     local_llm_model: str = "gpt-oss"
+    api_key: str = ""
     chat_temperature: float = 1.0
     code_temperature: float = 0.3
     n_parallel: int = 4