feat: integrate durable chat voice transport and deployment assets

agent/agent.py

@@ -13,9 +13,9 @@ from livekit import rtc
 from pathlib import Path
 
 try:
-    from agent.hermes_llm import HermesLLM
+    from agent.conversation_llm import ConversationLLM
 except ImportError:
-    from hermes_llm import HermesLLM
+    from conversation_llm import ConversationLLM
 
 try:
     from web.voice_catalog import normalize_default_voice
@@ -124,9 +124,9 @@ async def entrypoint(ctx: JobContext):
             raise RuntimeError("Account mismatch")
         return current["model_settings"]
 
-    bridge = HermesLLM(base_url=account["hermes_url"], api_key=account["hermes_api_key"], user_id=user_id,
-                       room_id=ctx.room.name, on_progress=progress, on_wait_notice=wait_notice,
-                       on_answer_start=stop_notice, get_model_settings=model_settings)
+    bridge = ConversationLLM(control_url=CONTROL_URL, secret=CONTROL_SECRET,
+                             room_id=ctx.room.name, on_progress=progress,
+                             on_wait_notice=wait_notice, on_answer_start=stop_notice)
     tts = azure.TTS(voice=DEFAULT_VOICE, sample_rate=24000,
                     speech_key=os.getenv("AZURE_SPEECH_KEY", ""), speech_region=os.getenv("AZURE_SPEECH_REGION", "eastus"))
     session = AgentSession(stt=azure.STT(speech_key=os.getenv("AZURE_SPEECH_KEY", ""),

agent/conversation_llm.py

@@ -0,0 +1,141 @@
+"""LiveKit speech adapter to the durable Hope coordinator, never a reasoning loop."""
+from __future__ import annotations
+import asyncio
+import json
+import uuid
+from dataclasses import replace
+import httpx
+from livekit.agents import llm
+from livekit.agents.types import DEFAULT_API_CONNECT_OPTIONS
+try:
+    from agent.hermes_llm import HermesLLMError, _sse
+except ImportError:
+    from hermes_llm import HermesLLMError, _sse
+
+
+class ConversationLLM(llm.LLM):
+    def __init__(self, *, control_url, secret, room_id, on_progress, on_wait_notice=None, on_answer_start=None):
+        super().__init__()
+        self.root = control_url.rstrip('/') + '/internal/chat/rooms/' + room_id
+        self.headers = {'X-Control-Secret': secret}
+        self.client = httpx.AsyncClient(follow_redirects=False, timeout=httpx.Timeout(660, connect=10))
+        self.on_progress, self.on_wait_notice, self.on_answer_start = on_progress, on_wait_notice, on_answer_start
+
+    @property
+    def model(self): return 'hermes-agent'
+    @property
+    def provider(self): return 'hermes'
+
+    def chat(self, *, chat_ctx, tools=None, conn_options=DEFAULT_API_CONNECT_OPTIONS, **kwargs):
+        return ConversationStream(self, chat_ctx, replace(conn_options, max_retry=0))
+
+    async def aclose(self):
+        await super().aclose()
+        await self.client.aclose()
+
+
+class ConversationStream(llm.LLMStream):
+    def __init__(self, bridge, chat_ctx, conn_options):
+        self.bridge = bridge
+        self.operation_id = next((m.id for m in reversed(chat_ctx.items) if isinstance(m, llm.ChatMessage) and m.role == 'user' and getattr(m, 'id', None)), uuid.uuid4().hex)
+        self.run_id = None
+        self.answer_started = False
+        self._close_task = None
+        super().__init__(bridge, chat_ctx=chat_ctx, tools=[], conn_options=conn_options)
+
+    async def aclose(self):
+        # LiveKit can close a stream from interruption and teardown concurrently.
+        # One underlying closer prevents a second cancellation from aborting
+        # the first closer's uncertain-admission reconciliation.
+        if self._close_task is None:
+            self._close_task = asyncio.create_task(super().aclose())
+        await asyncio.shield(self._close_task)
+
+    async def callback(self, callback, *args):
+        if callback:
+            try: await asyncio.wait_for(callback(*args), 5)
+            except Exception: pass  # feedback failure never repeats a tool action
+
+    async def progress(self, status, message):
+        await self.callback(self.bridge.on_progress, {'type': 'assistant_progress', 'request_id': self.run_id or self.operation_id, 'status': status, 'message': message})
+
+    async def _run(self):
+        user = next((m for m in reversed(self.chat_ctx.items) if isinstance(m, llm.ChatMessage) and m.role == 'user'), None)
+        text = '\n'.join(p for p in user.content if isinstance(p, str)) if user else ''
+        if not text.strip(): raise HermesLLMError('empty_voice_turn')
+        timer = None
+        admission = None
+        try:
+            # Do not abandon uncertain admission on barge-in. A successfully
+            # admitted remote action must be identified and cancelled, not left
+            # running just because the speech request lost its response.
+            admission = asyncio.create_task(self.bridge.client.post(self.bridge.root + '/messages', headers=self.bridge.headers,
+                json={'content': text, 'idempotency_key': self.operation_id, 'modality': 'voice'}, timeout=httpx.Timeout(20, connect=5)))
+            response = await asyncio.shield(admission)
+            if response.status_code != 202: raise HermesLLMError('voice_admission_failed')
+            self.run_id = response.json()['run']['id']
+            await self.progress('working', 'Working on your request.')
+            timer = asyncio.create_task(self.notice())
+            # Voice streams only completed public content. This avoids reading
+            # a draft that the backend subsequently replaces or rejects.
+            final = None
+            async with self.bridge.client.stream('GET', self.bridge.root + '/runs/' + self.run_id + '/events', headers=self.bridge.headers) as stream:
+                if stream.status_code != 200: raise HermesLLMError('voice_stream_failed')
+                async for event, raw in _sse(stream):
+                    item = json.loads(raw)
+                    if event == 'progress': await self.progress('working', 'Working on your request.')
+                    if event == 'message':
+                        message = item.get('message', item)
+                        if message.get('role') == 'assistant' and message.get('status') == 'completed': final = message.get('content', '')
+                    if event in ('run', 'done'):
+                        run = item.get('run', item)
+                        if run.get('status') in ('failed', 'cancelled'): raise HermesLLMError('voice_run_' + run['status'])
+                        if run.get('status') == 'awaiting_approval': await self.progress('working', 'Approval needed. Open this conversation to review it.')
+                        if event == 'done': break
+            if final:
+                self.answer_started = True
+                await self.callback(self.bridge.on_answer_start)
+                self._event_ch.send_nowait(llm.ChatChunk(id=self.operation_id, delta=llm.ChoiceDelta(role='assistant', content=final)))
+            else: raise HermesLLMError('empty_voice_answer')
+            await self.progress('completed', 'Completed.')
+        except asyncio.CancelledError:
+            # Persist cancellation by operation identity even when admission's
+            # response was lost. The coordinator tombstones the key, so a later
+            # identical admission cannot start an uncancelled remote action.
+            try:
+                await asyncio.shield(self.bridge.client.post(self.bridge.root + '/operations/' + self.operation_id + '/cancel',
+                    headers=self.bridge.headers, timeout=httpx.Timeout(10, connect=5)))
+            except Exception: pass
+            if not self.run_id and admission:
+                try:
+                    admitted = await asyncio.shield(admission)
+                    if admitted.status_code == 202:
+                        self.run_id = admitted.json()['run']['id']
+                except Exception:
+                    # A transport failure can be an uncertain acceptance. Retry
+                    # only the identical stable key; the coordinator deduplicates
+                    # admission, then return the run identity for cancellation.
+                    try:
+                        admitted = await self.bridge.client.post(self.bridge.root + '/messages', headers=self.bridge.headers,
+                            json={'content': text, 'idempotency_key': self.operation_id, 'modality': 'voice'})
+                        if admitted.status_code == 202:
+                            self.run_id = admitted.json()['run']['id']
+                    except Exception: pass
+            if self.run_id:
+                try:
+                    await asyncio.shield(self.bridge.client.post(self.bridge.root + '/runs/' + self.run_id + '/cancel', headers=self.bridge.headers))
+                except Exception: pass
+            await self.progress('failed', 'Request cancelled.')
+            raise
+        except Exception:
+            await self.progress('failed', 'Hope could not complete this response.')
+            raise HermesLLMError('voice_coordinator_failed') from None
+        finally:
+            if timer:
+                timer.cancel()
+                await asyncio.gather(timer, return_exceptions=True)
+
+    async def notice(self):
+        await asyncio.sleep(12)
+        if not self.answer_started:
+            await self.callback(self.bridge.on_wait_notice, "I'm still working on that. I'll let you know when it's ready.")