solana-agent 31.2.2__tar.gz → 31.2.4__tar.gz

{solana_agent-31.2.2 → solana_agent-31.2.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: solana-agent
-Version: 31.2.2
+Version: 31.2.4
 Summary: AI Agents for Solana
 License: MIT
 Keywords: solana,solana ai,solana agent,ai,ai agent,ai agents
@@ -317,27 +317,39 @@ Due to the overhead of the router (API call) - realtime only supports a single a
 
 Realtime uses MongoDB for memory so Zep is not needed.
 
+This example will work using expo-audio on Android and iOS.
+
 ```python
 from solana_agent import SolanaAgent
 
 solana_agent = SolanaAgent(config=config)
 
-# Example: mobile sends MP4/AAC; server encodes output to AAC
-audio_content = await audio_file.read()  # bytes
-async for audio_chunk in solana_agent.process(
-    "user123",                    # required
-    audio_content,                # required
-    realtime=True,                # optional (default False)
-    output_format="audio",        # required
-    vad=True,                     # enable VAD (optional)
-    rt_encode_input=True,         # accept compressed input (optional)
-    rt_encode_output=True,        # encode output for client (optional)
-    rt_voice="marin"              # the voice to use for interactions (optional)
-    audio_input_format="mp4",     # client transport (optional)
-    audio_output_format="aac"     # client transport (optional)
-):
-    handle_audio(audio_chunk)
-```
+audio_content = await audio_file.read()
+
+async def generate():
+    async for chunk in solana_agent.process(
+        user_id=user_id, 
+        message=audio_content,
+        realtime=True,
+        rt_encode_input=True,
+        rt_encode_output=True,
+        rt_voice="marin",
+        output_format="audio",
+        audio_output_format="mp3",
+        audio_input_format="mp4",
+    ):
+        yield chunk
+
+return StreamingResponse(
+    content=generate(),
+    media_type="audio/mp3",
+    headers={
+        "Cache-Control": "no-store",
+        "Pragma": "no-cache",
+        "Content-Disposition": "inline; filename=stream.mp3",
+        "X-Accel-Buffering": "no",
+    },
+)
 
 ### Image/Text Streaming
 

{solana_agent-31.2.2 → solana_agent-31.2.4}/README.md

@@ -281,27 +281,39 @@ Due to the overhead of the router (API call) - realtime only supports a single a
 
 Realtime uses MongoDB for memory so Zep is not needed.
 
+This example will work using expo-audio on Android and iOS.
+
 ```python
 from solana_agent import SolanaAgent
 
 solana_agent = SolanaAgent(config=config)
 
-# Example: mobile sends MP4/AAC; server encodes output to AAC
-audio_content = await audio_file.read()  # bytes
-async for audio_chunk in solana_agent.process(
-    "user123",                    # required
-    audio_content,                # required
-    realtime=True,                # optional (default False)
-    output_format="audio",        # required
-    vad=True,                     # enable VAD (optional)
-    rt_encode_input=True,         # accept compressed input (optional)
-    rt_encode_output=True,        # encode output for client (optional)
-    rt_voice="marin"              # the voice to use for interactions (optional)
-    audio_input_format="mp4",     # client transport (optional)
-    audio_output_format="aac"     # client transport (optional)
-):
-    handle_audio(audio_chunk)
-```
+audio_content = await audio_file.read()
+
+async def generate():
+    async for chunk in solana_agent.process(
+        user_id=user_id, 
+        message=audio_content,
+        realtime=True,
+        rt_encode_input=True,
+        rt_encode_output=True,
+        rt_voice="marin",
+        output_format="audio",
+        audio_output_format="mp3",
+        audio_input_format="mp4",
+    ):
+        yield chunk
+
+return StreamingResponse(
+    content=generate(),
+    media_type="audio/mp3",
+    headers={
+        "Cache-Control": "no-store",
+        "Pragma": "no-cache",
+        "Content-Disposition": "inline; filename=stream.mp3",
+        "X-Accel-Buffering": "no",
+    },
+)
 
 ### Image/Text Streaming
 

{solana_agent-31.2.2 → solana_agent-31.2.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "solana-agent"
-version = "31.2.2"
+version = "31.2.4"
 description = "AI Agents for Solana"
 authors = ["Bevan Hunt <bevan@bevanhunt.com>"]
 license = "MIT"

{solana_agent-31.2.2 → solana_agent-31.2.4}/solana_agent/adapters/ffmpeg_transcoder.py

@@ -4,6 +4,8 @@ import asyncio
 import contextlib
 import logging
 from typing import List, AsyncGenerator
+import tempfile
+import os
 
 from solana_agent.interfaces.providers.audio import AudioTranscoder
 
@@ -49,11 +51,45 @@ class FFmpegTranscoder(AudioTranscoder):
             rate_hz,
             len(audio_bytes),
         )
-        # Prefer to hint format for common containers/codecs; ffmpeg can still autodetect if hint is wrong.
-        hinted_format = None
+        # iOS-recorded MP4/M4A often requires a seekable input for reliable demuxing.
+        # Decode from a temporary file instead of stdin for MP4/M4A.
         if input_mime in ("audio/mp4", "audio/m4a"):
-            hinted_format = "mp4"
-        elif input_mime in ("audio/aac",):
+            suffix = ".m4a" if input_mime == "audio/m4a" else ".mp4"
+            tmp_path = None
+            try:
+                with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tf:
+                    tmp_path = tf.name
+                    tf.write(audio_bytes)
+                args = [
+                    "-hide_banner",
+                    "-loglevel",
+                    "error",
+                    "-i",
+                    tmp_path,
+                    "-vn",  # ignore any video tracks
+                    "-acodec",
+                    "pcm_s16le",
+                    "-ac",
+                    "1",
+                    "-ar",
+                    str(rate_hz),
+                    "-f",
+                    "s16le",
+                    "pipe:1",
+                ]
+                out = await self._run_ffmpeg(args, b"")
+                logger.info(
+                    "Transcoded (MP4/M4A temp-file) to PCM16: output_len=%d", len(out)
+                )
+                return out
+            finally:
+                if tmp_path:
+                    with contextlib.suppress(Exception):
+                        os.remove(tmp_path)
+
+        # For other formats, prefer a format hint when helpful and decode from stdin.
+        hinted_format = None
+        if input_mime in ("audio/aac",):
             # Raw AAC is typically in ADTS stream format
             hinted_format = "adts"
         elif input_mime in ("audio/ogg", "audio/webm"):
@@ -88,13 +124,14 @@ class FFmpegTranscoder(AudioTranscoder):
     async def from_pcm16(  # pragma: no cover
         self, pcm16_bytes: bytes, output_mime: str, rate_hz: int
     ) -> bytes:
-        """Encode PCM16LE to desired format (currently AAC ADTS for mobile streaming)."""
+        """Encode PCM16LE to desired format (AAC ADTS, fragmented MP4, or MP3)."""
         logger.info(
             "Encode from PCM16: output_mime=%s, rate_hz=%d, input_len=%d",
             output_mime,
             rate_hz,
             len(pcm16_bytes),
         )
+
         if output_mime in ("audio/mpeg", "audio/mp3"):
             # Encode to MP3 (often better streaming compatibility on mobile)
             args = [
@@ -122,8 +159,9 @@ class FFmpegTranscoder(AudioTranscoder):
                 "Encoded from PCM16 to %s: output_len=%d", output_mime, len(out)
             )
             return out
-        if output_mime in ("audio/aac", "audio/mp4", "audio/m4a"):
-            # Encode to AAC in ADTS stream; clients can play it as AAC.
+
+        if output_mime in ("audio/aac",):
+            # Encode to AAC in ADTS stream; good for streaming over sockets/HTTP chunked
             args = [
                 "-hide_banner",
                 "-loglevel",
@@ -149,6 +187,38 @@ class FFmpegTranscoder(AudioTranscoder):
                 "Encoded from PCM16 to %s: output_len=%d", output_mime, len(out)
             )
             return out
+
+        if output_mime in ("audio/mp4", "audio/m4a"):
+            # Encode to fragmented MP4 (fMP4) with AAC for better iOS compatibility
+            # For streaming, write an initial moov and fragment over stdout.
+            args = [
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-f",
+                "s16le",
+                "-ac",
+                "1",
+                "-ar",
+                str(rate_hz),
+                "-i",
+                "pipe:0",
+                "-c:a",
+                "aac",
+                "-b:a",
+                "96k",
+                "-movflags",
+                "+frag_keyframe+empty_moov",
+                "-f",
+                "mp4",
+                "pipe:1",
+            ]
+            out = await self._run_ffmpeg(args, pcm16_bytes)
+            logger.info(
+                "Encoded from PCM16 to %s (fMP4): output_len=%d", output_mime, len(out)
+            )
+            return out
+
         # Default: passthrough
         logger.info("Encode passthrough (no change), output_len=%d", len(pcm16_bytes))
         return pcm16_bytes
@@ -187,7 +257,7 @@ class FFmpegTranscoder(AudioTranscoder):
                 "mp3",
                 "pipe:1",
             ]
-        elif output_mime in ("audio/aac", "audio/mp4", "audio/m4a"):
+        elif output_mime in ("audio/aac",):
             args = [
                 "-hide_banner",
                 "-loglevel",
@@ -208,6 +278,29 @@ class FFmpegTranscoder(AudioTranscoder):
                 "adts",
                 "pipe:1",
             ]
+        elif output_mime in ("audio/mp4", "audio/m4a"):
+            args = [
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-f",
+                "s16le",
+                "-ac",
+                "1",
+                "-ar",
+                str(rate_hz),
+                "-i",
+                "pipe:0",
+                "-c:a",
+                "aac",
+                "-b:a",
+                "96k",
+                "-movflags",
+                "+frag_keyframe+empty_moov",
+                "-f",
+                "mp4",
+                "pipe:1",
+            ]
         else:
             # Passthrough streaming: just yield input
             async for chunk in pcm_iter:

{solana_agent-31.2.2 → solana_agent-31.2.4}/solana_agent/adapters/openai_realtime_ws.py

@@ -325,7 +325,26 @@ class OpenAIRealtimeWebSocketSession(BaseRealtimeSession):
                             try:
                                 chunk = base64.b64decode(b64)
                                 self._audio_queue.put_nowait(chunk)
-                                logger.info("Audio delta bytes=%d", len(chunk))
+                                # Ownership/response tagging for diagnostics
+                                try:
+                                    owner = getattr(self, "_owner_user_id", None)
+                                except Exception:
+                                    owner = None
+                                try:
+                                    rid = getattr(self, "_active_response_id", None)
+                                except Exception:
+                                    rid = None
+                                try:
+                                    gen = int(getattr(self, "_response_generation", 0))
+                                except Exception:
+                                    gen = None
+                                logger.info(
+                                    "Audio delta bytes=%d owner=%s rid=%s gen=%s",
+                                    len(chunk),
+                                    owner,
+                                    rid,
+                                    gen,
+                                )
                                 try:
                                     # New response detected if we were previously inactive
                                     if not getattr(self, "_response_active", False):
@@ -492,8 +511,25 @@ class OpenAIRealtimeWebSocketSession(BaseRealtimeSession):
                         "response.audio.done",
                     ):
                         # End of audio stream for the response; stop audio iterator but keep WS open for transcripts
+                        try:
+                            owner = getattr(self, "_owner_user_id", None)
+                        except Exception:
+                            owner = None
+                        try:
+                            rid = (data.get("response") or {}).get("id") or getattr(
+                                self, "_active_response_id", None
+                            )
+                        except Exception:
+                            rid = None
+                        try:
+                            gen = int(getattr(self, "_response_generation", 0))
+                        except Exception:
+                            gen = None
                         logger.info(
-                            "Realtime WS: output audio done; ending audio stream"
+                            "Realtime WS: output audio done; owner=%s rid=%s gen=%s",
+                            owner,
+                            rid,
+                            gen,
                         )
                         # If we have a buffered transcript for this response, flush it now
                         try:
@@ -1001,6 +1037,15 @@ class OpenAIRealtimeWebSocketSession(BaseRealtimeSession):
         if "tools" in patch:
             patch["tools"] = _strip_tool_strict(patch["tools"])  # idempotent
 
+        # Per server requirements, always include session.type and output_modalities
+        try:
+            patch["type"] = "realtime"
+            # Preserve caller-provided output_modalities if present, otherwise default to audio
+            if "output_modalities" not in patch:
+                patch["output_modalities"] = ["audio"]
+        except Exception:
+            pass
+
         payload = {"type": "session.update", "session": patch}
         # Mark awaiting updated and store last patch
         self._last_session_patch = patch or {}

{solana_agent-31.2.2 → solana_agent-31.2.4}/solana_agent/services/query.py

@@ -67,10 +67,117 @@ class QueryService(QueryServiceInterface):
         self._sticky_sessions: Dict[str, Dict[str, Any]] = {}
         # Optional realtime service attached by factory (populated in factory)
         self.realtime = None  # type: ignore[attr-defined]
-        # Persistent realtime WS per user for push-to-talk reuse
-        self._rt_services = {}
+        # Persistent realtime WS pool per user for reuse across turns/devices
+        # { user_id: [RealtimeService, ...] }
+        self._rt_services: Dict[str, List[Any]] = {}
+        # Global lock for creating/finding per-user sessions
         self._rt_lock = asyncio.Lock()
 
+    async def _try_acquire_lock(self, lock: asyncio.Lock) -> bool:
+        try:
+            await asyncio.wait_for(lock.acquire(), timeout=0)
+            return True
+        except asyncio.TimeoutError:
+            return False
+        except Exception:
+            return False
+
+    async def _alloc_realtime_session(
+        self,
+        user_id: str,
+        *,
+        api_key: str,
+        rt_voice: str,
+        final_instructions: str,
+        initial_tools: Optional[List[Dict[str, Any]]],
+        encode_in: bool,
+        encode_out: bool,
+        audio_input_format: str,
+        audio_output_format: str,
+    ) -> Any:
+        """Get a free (or new) realtime session for this user. Marks it busy via an internal lock.
+
+        Returns the RealtimeService with an acquired _in_use_lock that MUST be released by caller.
+        """
+        from solana_agent.interfaces.providers.realtime import (
+            RealtimeSessionOptions,
+        )
+        from solana_agent.adapters.openai_realtime_ws import (
+            OpenAIRealtimeWebSocketSession,
+        )
+        from solana_agent.adapters.ffmpeg_transcoder import FFmpegTranscoder
+
+        def _mime_from(fmt: str) -> str:
+            f = (fmt or "").lower()
+            return {
+                "aac": "audio/aac",
+                "mp3": "audio/mpeg",
+                "mp4": "audio/mp4",
+                "m4a": "audio/mp4",
+                "mpeg": "audio/mpeg",
+                "mpga": "audio/mpeg",
+                "wav": "audio/wav",
+                "flac": "audio/flac",
+                "opus": "audio/opus",
+                "ogg": "audio/ogg",
+                "webm": "audio/webm",
+                "pcm": "audio/pcm",
+            }.get(f, "audio/pcm")
+
+        async with self._rt_lock:
+            pool = self._rt_services.get(user_id) or []
+            # Try to reuse an idle session strictly owned by this user
+            for rt in pool:
+                # Extra safety: never reuse a session from another user
+                owner = getattr(rt, "_owner_user_id", None)
+                if owner is not None and owner != user_id:
+                    continue
+                lock = getattr(rt, "_in_use_lock", None)
+                if lock is None:
+                    lock = asyncio.Lock()
+                    setattr(rt, "_in_use_lock", lock)
+                if not lock.locked():
+                    if await self._try_acquire_lock(lock):
+                        return rt
+            # None free: create a new session
+            opts = RealtimeSessionOptions(
+                model="gpt-realtime",
+                voice=rt_voice,
+                vad_enabled=False,
+                input_rate_hz=24000,
+                output_rate_hz=24000,
+                input_mime="audio/pcm",
+                output_mime="audio/pcm",
+                tools=initial_tools or None,
+                tool_choice="auto",
+            )
+            try:
+                opts.instructions = final_instructions
+                opts.voice = rt_voice
+            except Exception:
+                pass
+            conv_session = OpenAIRealtimeWebSocketSession(api_key=api_key, options=opts)
+            transcoder = FFmpegTranscoder() if (encode_in or encode_out) else None
+            from solana_agent.services.realtime import RealtimeService
+
+            rt = RealtimeService(
+                session=conv_session,
+                options=opts,
+                transcoder=transcoder,
+                accept_compressed_input=encode_in,
+                client_input_mime=_mime_from(audio_input_format),
+                encode_output=encode_out,
+                client_output_mime=_mime_from(audio_output_format),
+            )
+            # Tag ownership to prevent any cross-user reuse
+            setattr(rt, "_owner_user_id", user_id)
+            setattr(rt, "_in_use_lock", asyncio.Lock())
+            # Mark busy
+            await getattr(rt, "_in_use_lock").acquire()
+            pool.append(rt)
+            self._rt_services[user_id] = pool
+            return rt
+
     def _get_sticky_agent(self, user_id: str) -> Optional[str]:
         sess = self._sticky_sessions.get(user_id)
         return sess.get("agent") if isinstance(sess, dict) else None
@@ -554,14 +661,7 @@ class QueryService(QueryServiceInterface):
                 final_instructions = "\n\n".join([p for p in parts if p])
 
                 # 4) Open a single WS session for assistant audio
-                from solana_agent.adapters.openai_realtime_ws import (
-                    OpenAIRealtimeWebSocketSession,
-                )
-                from solana_agent.interfaces.providers.realtime import (
-                    RealtimeSessionOptions,
-                )
-                from solana_agent.services.realtime import RealtimeService
-                from solana_agent.adapters.ffmpeg_transcoder import FFmpegTranscoder
+                # Realtime imports handled inside allocator helper
 
                 api_key = None
                 try:
@@ -600,171 +700,160 @@ class QueryService(QueryServiceInterface):
                     or (is_audio_bytes and audio_input_format.lower() != "pcm")
                 )
 
-                async with self._rt_lock:
-                    rt = self._rt_services.get(user_id)
-                    if not rt or not isinstance(rt, RealtimeService):
-                        opts = RealtimeSessionOptions(
-                            model="gpt-realtime",
-                            voice=rt_voice,
-                            vad_enabled=False,  # no input audio
-                            input_rate_hz=24000,
-                            output_rate_hz=24000,
-                            input_mime="audio/pcm",
-                            output_mime="audio/pcm",
-                            tools=initial_tools or None,
-                            tool_choice="auto",
-                        )
-                        # Ensure initial session.update carries instructions/voice
+                # Allocate or reuse a realtime session for this specific request/user
+                rt = await self._alloc_realtime_session(
+                    user_id,
+                    api_key=api_key,
+                    rt_voice=rt_voice,
+                    final_instructions=final_instructions,
+                    initial_tools=initial_tools,
+                    encode_in=encode_in,
+                    encode_out=encode_out,
+                    audio_input_format=audio_input_format,
+                    audio_output_format=audio_output_format,
+                )
+                # Ensure lock is released no matter what
+                try:
+                    # Tool executor
+                    async def _exec(
+                        tool_name: str, args: Dict[str, Any]
+                    ) -> Dict[str, Any]:
                         try:
-                            opts.instructions = final_instructions
-                            opts.voice = rt_voice
-                        except Exception:
-                            pass
-                        conv_session = OpenAIRealtimeWebSocketSession(
-                            api_key=api_key, options=opts
-                        )
-                        transcoder = (
-                            FFmpegTranscoder() if (encode_in or encode_out) else None
-                        )
-                        rt = RealtimeService(
-                            session=conv_session,
-                            options=opts,
-                            transcoder=transcoder,
-                            accept_compressed_input=encode_in,
-                            client_input_mime=_mime_from(audio_input_format),
-                            encode_output=encode_out,
-                            client_output_mime=_mime_from(audio_output_format),
-                        )
-                        self._rt_services[user_id] = rt
+                            return await self.agent_service.execute_tool(
+                                agent_name, tool_name, args or {}
+                            )
+                        except Exception as e:
+                            return {"status": "error", "message": str(e)}
 
-                # Tool executor
-                async def _exec(tool_name: str, args: Dict[str, Any]) -> Dict[str, Any]:
+                    # If possible, set on underlying session
                     try:
-                        return await self.agent_service.execute_tool(
-                            agent_name, tool_name, args or {}
-                        )
-                    except Exception as e:
-                        return {"status": "error", "message": str(e)}
-
-                # If possible, set on underlying session
-                try:
-                    if hasattr(rt, "_session"):
-                        getattr(rt, "_session").set_tool_executor(_exec)  # type: ignore[attr-defined]
-                except Exception:
-                    pass
-
-                # Connect/configure
-                if not getattr(rt, "_connected", False):
-                    await rt.start()
-                await rt.configure(
-                    voice=rt_voice,
-                    vad_enabled=bool(vad) if vad is not None else False,
-                    instructions=final_instructions,
-                    tools=initial_tools or None,
-                    tool_choice="auto",
-                )
+                        if hasattr(rt, "_session"):
+                            getattr(rt, "_session").set_tool_executor(_exec)  # type: ignore[attr-defined]
+                    except Exception:
+                        pass
 
-                # Ensure clean input buffers for this turn
-                try:
-                    await rt.clear_input()
-                except Exception:
-                    pass
-                # Also reset any leftover output audio so new turn doesn't replay old chunks
-                try:
-                    if hasattr(rt, "reset_output_stream"):
-                        rt.reset_output_stream()
-                except Exception:
-                    pass
+                    # Connect/configure
+                    if not getattr(rt, "_connected", False):
+                        await rt.start()
+                    await rt.configure(
+                        voice=rt_voice,
+                        vad_enabled=bool(vad) if vad is not None else False,
+                        instructions=final_instructions,
+                        tools=initial_tools or None,
+                        tool_choice="auto",
+                    )
 
-                # Persist once per turn
-                turn_id = await self.realtime_begin_turn(user_id)
-                if turn_id and user_text:
+                    # Ensure clean input buffers for this turn
+                    try:
+                        await rt.clear_input()
+                    except Exception:
+                        pass
+                    # Also reset any leftover output audio so new turn doesn't replay old chunks
                     try:
-                        await self.realtime_update_user(user_id, turn_id, user_text)
+                        if hasattr(rt, "reset_output_stream"):
+                            rt.reset_output_stream()
                     except Exception:
                         pass
 
-                # Feed audio into WS if audio bytes provided; else use input_text
-                if is_audio_bytes:
-                    bq = bytes(query)
-                    logger.info(
-                        "Realtime: appending input audio to WS via FFmpeg, len=%d, fmt=%s",
-                        len(bq),
-                        audio_input_format,
-                    )
-                    await rt.append_audio(bq)
-                    vad_enabled_value = bool(vad) if vad is not None else False
-                    if not vad_enabled_value:
-                        await rt.commit_input()
-                        # Manually trigger response when VAD is disabled
-                        await rt.create_response({})
+                    # Persist once per turn
+                    turn_id = await self.realtime_begin_turn(user_id)
+                    if turn_id and user_text:
+                        try:
+                            await self.realtime_update_user(user_id, turn_id, user_text)
+                        except Exception:
+                            pass
+
+                    # Feed audio into WS if audio bytes provided; else use input_text
+                    if is_audio_bytes:
+                        bq = bytes(query)
+                        logger.info(
+                            "Realtime: appending input audio to WS via FFmpeg, len=%d, fmt=%s",
+                            len(bq),
+                            audio_input_format,
+                        )
+                        await rt.append_audio(bq)
+                        vad_enabled_value = bool(vad) if vad is not None else False
+                        if not vad_enabled_value:
+                            await rt.commit_input()
+                            # Manually trigger response when VAD is disabled
+                            await rt.create_response({})
+                        else:
+                            # With server VAD enabled, the model will auto-create a response at end of speech
+                            logger.debug(
+                                "Realtime: VAD enabled — skipping manual response.create"
+                            )
                     else:
-                        # With server VAD enabled, the model will auto-create a response at end of speech
-                        logger.debug(
-                            "Realtime: VAD enabled — skipping manual response.create"
+                        # Rely on configured session voice; attach input_text only
+                        await rt.create_response(
+                            {
+                                "modalities": ["audio"],
+                                "input": [
+                                    {"type": "input_text", "text": user_text or ""}
+                                ],
+                            }
                         )
-                else:
-                    # Rely on configured session voice; attach input_text only
-                    await rt.create_response(
-                        {
-                            "modalities": ["audio"],
-                            "input": [{"type": "input_text", "text": user_text or ""}],
-                        }
-                    )
 
-                # Collect audio and transcripts
-                user_tr = ""
-                asst_tr = ""
+                    # Collect audio and transcripts
+                    user_tr = ""
+                    asst_tr = ""
 
-                async def _drain_in_tr():
-                    nonlocal user_tr
-                    async for t in rt.iter_input_transcript():
-                        if t:
-                            user_tr += t
+                    async def _drain_in_tr():
+                        nonlocal user_tr
+                        async for t in rt.iter_input_transcript():
+                            if t:
+                                user_tr += t
 
-                async def _drain_out_tr():
-                    nonlocal asst_tr
-                    async for t in rt.iter_output_transcript():
-                        if t:
-                            asst_tr += t
+                    async def _drain_out_tr():
+                        nonlocal asst_tr
+                        async for t in rt.iter_output_transcript():
+                            if t:
+                                asst_tr += t
 
-                in_task = asyncio.create_task(_drain_in_tr())
-                out_task = asyncio.create_task(_drain_out_tr())
-                try:
-                    async for audio_chunk in rt.iter_output_audio_encoded():
-                        yield audio_chunk
-                finally:
-                    in_task.cancel()
-                    out_task.cancel()
-                    # If no WS input transcript was captured, fall back to HTTP STT result
-                    if not user_tr:
-                        try:
-                            if "stt_task" in locals() and stt_task is not None:
-                                user_tr = await stt_task
-                        except Exception:
-                            pass
-                    if turn_id:
-                        try:
-                            if user_tr:
-                                await self.realtime_update_user(
-                                    user_id, turn_id, user_tr
-                                )
-                            if asst_tr:
-                                await self.realtime_update_assistant(
-                                    user_id, turn_id, asst_tr
-                                )
-                        except Exception:
-                            pass
+                    in_task = asyncio.create_task(_drain_in_tr())
+                    out_task = asyncio.create_task(_drain_out_tr())
+                    try:
+                        async for audio_chunk in rt.iter_output_audio_encoded():
+                            yield audio_chunk
+                    finally:
+                        in_task.cancel()
+                        out_task.cancel()
+                        # If no WS input transcript was captured, fall back to HTTP STT result
+                        if not user_tr:
+                            try:
+                                if "stt_task" in locals() and stt_task is not None:
+                                    user_tr = await stt_task
+                            except Exception:
+                                pass
+                        if turn_id:
+                            try:
+                                if user_tr:
+                                    await self.realtime_update_user(
+                                        user_id, turn_id, user_tr
+                                    )
+                                if asst_tr:
+                                    await self.realtime_update_assistant(
+                                        user_id, turn_id, asst_tr
+                                    )
+                            except Exception:
+                                pass
+                            try:
+                                await self.realtime_finalize_turn(user_id, turn_id)
+                            except Exception:
+                                pass
+                        # Clear input buffer for next turn reuse
                         try:
-                            await self.realtime_finalize_turn(user_id, turn_id)
+                            await rt.clear_input()
                         except Exception:
                             pass
-                    # Clear input buffer for next turn reuse
+                finally:
+                    # Always release the session for reuse by other concurrent requests/devices
                     try:
-                        await rt.clear_input()
+                        lock = getattr(rt, "_in_use_lock", None)
+                        if lock and lock.locked():
+                            lock.release()
                     except Exception:
                         pass
-                return
+                    return
 
             # 1) Transcribe audio or accept text
             user_text = ""