openrtc 0.1.0__tar.gz → 0.2.2__tar.gz

{openrtc-0.1.0 → openrtc-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: openrtc
-Version: 0.1.0
+Version: 0.2.2
 Summary: Run multiple LiveKit voice agents in a single shared worker process.
 Project-URL: Homepage, https://github.com/mahimailabs/openrtc
 Project-URL: Repository, https://github.com/mahimailabs/openrtc
@@ -20,6 +20,7 @@ Classifier: Topic :: Multimedia :: Sound/Audio
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: <3.14,>=3.11
 Requires-Dist: livekit-agents[openai,silero,turn-detector]~=1.5
+Requires-Dist: watchfiles<2,>=0.21
 Provides-Extra: cli
 Requires-Dist: rich>=13; extra == 'cli'
 Requires-Dist: typer>=0.12; extra == 'cli'
@@ -265,6 +266,71 @@ footprint. Validate against the §8.4 real-LiveKit integration test
 `OPENAI_API_KEY`) before quoting a per-session memory number to your
 operators.
 
+### Throughput: steady-state event-loop p99
+
+Memory density is only half the question. N sessions share one event loop and
+one GIL, so the other half is whether the loop keeps up.
+`tests/benchmarks/throughput.py` drives N concurrent sessions through the real
+Silero VAD over synthetic 16 kHz PCM at 50 fps (the continuous on-loop CPU cost)
+and measures event-loop p99 latency, separating the startup burst from steady
+state.
+
+```bash
+uv run python tests/benchmarks/throughput.py --sessions 1,10,25,50,100
+```
+
+Sample sweep (Apple M-series laptop, `vad` workload, steady state):
+
+| Sessions | steady-state loop p99 | peak RSS |
+| ---: | ---: | ---: |
+|   1 | 0.9 ms | 160 MB |
+|  10 | 1.3 ms | 160 MB |
+|  25 | 1.2 ms | 160 MB |
+|  50 | 1.1 ms | 160 MB |
+| 100 | 2.8 ms | 160 MB |
+
+Steady-state VAD inference stays well under a 100 ms loop-latency budget to 100
+sessions, with flat resident memory (the model loads once). The expensive,
+bursty part is session *startup* (each `session.start()` plus greeting), which
+the benchmark reports as a separate `startup_p99` column and which dominates
+early-life latency. This workload models the continuous VAD path, not the full
+STT/LLM/TTS orchestration, so read it as the on-loop-CPU ceiling rather than a
+full-pipeline guarantee. Run it on your own hardware before quoting a
+sessions-per-worker number.
+
+### Prove it on your machine
+
+The process column above is estimated. This script measures both models for
+real on your laptop: it spawns one subprocess per session for the
+process-per-session model, runs the same number of sessions as `asyncio`
+tasks in a single process for the coroutine model, then prints the memory
+used each way. No LiveKit server, no API keys, no model download.
+
+```bash
+uv run python examples/density_demo.py                 # 16 sessions
+uv run python examples/density_demo.py --sessions 32   # the gap widens with N
+uv run python examples/density_demo.py --sessions 50 --load-vad   # adds the shared Silero VAD model
+```
+
+Sample output (Apple M-series laptop, import-only mode):
+
+```text
+Hosting 16 concurrent voice sessions. Measuring resident memory.
+
+  livekit-agents (process per session):     1861 MB total   ( 116.3 MB/session)
+  OpenRTC coroutine pool (one process):      195 MB total   (  12.2 MB/session)
+
+  OpenRTC uses 9.5x less memory for the same 16 sessions.
+```
+
+Your numbers vary by machine, and the ratio grows as you raise `--sessions`
+(the coroutine pool pays the import cost once and amortizes it across every
+session). This default mode counts only the `livekit-agents` import cost, so
+it is a conservative lower bound: `--load-vad` adds the shared Silero VAD
+model weights (paid once in the pool, once per process otherwise), and
+`tests/benchmarks/density.py --sessions 50` proves the 50-sessions-under-4-GB
+ceiling. The full script is [examples/density_demo.py](examples/density_demo.py).
+
 ## Routing
 
 One process hosts several agent classes, so each session must resolve to a single registered name. `AgentPool` resolves the agent in this order:

{openrtc-0.1.0 → openrtc-0.2.2}/README.md

@@ -234,6 +234,71 @@ footprint. Validate against the §8.4 real-LiveKit integration test
 `OPENAI_API_KEY`) before quoting a per-session memory number to your
 operators.
 
+### Throughput: steady-state event-loop p99
+
+Memory density is only half the question. N sessions share one event loop and
+one GIL, so the other half is whether the loop keeps up.
+`tests/benchmarks/throughput.py` drives N concurrent sessions through the real
+Silero VAD over synthetic 16 kHz PCM at 50 fps (the continuous on-loop CPU cost)
+and measures event-loop p99 latency, separating the startup burst from steady
+state.
+
+```bash
+uv run python tests/benchmarks/throughput.py --sessions 1,10,25,50,100
+```
+
+Sample sweep (Apple M-series laptop, `vad` workload, steady state):
+
+| Sessions | steady-state loop p99 | peak RSS |
+| ---: | ---: | ---: |
+|   1 | 0.9 ms | 160 MB |
+|  10 | 1.3 ms | 160 MB |
+|  25 | 1.2 ms | 160 MB |
+|  50 | 1.1 ms | 160 MB |
+| 100 | 2.8 ms | 160 MB |
+
+Steady-state VAD inference stays well under a 100 ms loop-latency budget to 100
+sessions, with flat resident memory (the model loads once). The expensive,
+bursty part is session *startup* (each `session.start()` plus greeting), which
+the benchmark reports as a separate `startup_p99` column and which dominates
+early-life latency. This workload models the continuous VAD path, not the full
+STT/LLM/TTS orchestration, so read it as the on-loop-CPU ceiling rather than a
+full-pipeline guarantee. Run it on your own hardware before quoting a
+sessions-per-worker number.
+
+### Prove it on your machine
+
+The process column above is estimated. This script measures both models for
+real on your laptop: it spawns one subprocess per session for the
+process-per-session model, runs the same number of sessions as `asyncio`
+tasks in a single process for the coroutine model, then prints the memory
+used each way. No LiveKit server, no API keys, no model download.
+
+```bash
+uv run python examples/density_demo.py                 # 16 sessions
+uv run python examples/density_demo.py --sessions 32   # the gap widens with N
+uv run python examples/density_demo.py --sessions 50 --load-vad   # adds the shared Silero VAD model
+```
+
+Sample output (Apple M-series laptop, import-only mode):
+
+```text
+Hosting 16 concurrent voice sessions. Measuring resident memory.
+
+  livekit-agents (process per session):     1861 MB total   ( 116.3 MB/session)
+  OpenRTC coroutine pool (one process):      195 MB total   (  12.2 MB/session)
+
+  OpenRTC uses 9.5x less memory for the same 16 sessions.
+```
+
+Your numbers vary by machine, and the ratio grows as you raise `--sessions`
+(the coroutine pool pays the import cost once and amortizes it across every
+session). This default mode counts only the `livekit-agents` import cost, so
+it is a conservative lower bound: `--load-vad` adds the shared Silero VAD
+model weights (paid once in the pool, once per process otherwise), and
+`tests/benchmarks/density.py --sessions 50` proves the 50-sessions-under-4-GB
+ceiling. The full script is [examples/density_demo.py](examples/density_demo.py).
+
 ## Routing
 
 One process hosts several agent classes, so each session must resolve to a single registered name. `AgentPool` resolves the agent in this order:

{openrtc-0.1.0 → openrtc-0.2.2}/docs/changelog.md

@@ -147,6 +147,30 @@ contributor onboarding matches what's in the repo.
 
 <!-- releases -->
 
+## [0.2.1] - 2026-05-06
+
+## What's Changed
+* [v0.2.1] File watcher infrastructure for agent code (MAH-80) by @mahimairaja in https://github.com/mahimailabs/openrtc-runtime/pull/39
+
+
+**Full Changelog**: https://github.com/mahimailabs/openrtc-runtime/compare/v0.1.0...v0.2.1
+
+---
+
+## [0.1.0] - 2026-05-06
+
+## What's Changed
+* Feat: light websocket by @mahimairaja in https://github.com/mahimailabs/openrtc-runtime/pull/30
+* docs: bring docs/ in sync with v0.1 surface by @mahimairaja in https://github.com/mahimailabs/openrtc-runtime/pull/35
+* Feat: struc refac by @mahimairaja in https://github.com/mahimailabs/openrtc-runtime/pull/36
+* Feat/coroutine pool by @mahimairaja in https://github.com/mahimailabs/openrtc-runtime/pull/37
+* Feat/coroutine pool prod by @mahimairaja in https://github.com/mahimailabs/openrtc-runtime/pull/38
+
+
+**Full Changelog**: https://github.com/mahimailabs/openrtc-runtime/compare/v0.0.17...v0.1.0
+
+---
+
 ## [0.0.17] - 2026-04-03
 
 ## What's Changed

openrtc-0.2.2/examples/density_demo.py

@@ -0,0 +1,163 @@
+"""Prove the OpenRTC density win, on one laptop, with real numbers.
+
+The claim: livekit-agents runs roughly one OS process per session (about
+3 GB each in production). OpenRTC's coroutine pool runs N sessions as
+asyncio tasks inside a single process, so the heavy per-process cost
+(Python interpreter, the livekit-agents import graph, and shared models
+like Silero VAD and the turn detector) is paid ONCE instead of N times.
+
+This script measures both models for real:
+
+  * "process-per-session" (what vanilla livekit-agents does):
+    spawn N subprocesses, each imports the agent stack and holds a
+    per-session buffer. We sum the resident memory across all of them.
+
+  * "OpenRTC coroutine pool" (the default isolation mode):
+    import the stack ONCE, run N asyncio sessions in this single process,
+    each holding the same per-session buffer. We read this process's
+    resident memory.
+
+Then it prints total memory each way, memory per session, and the ratio.
+No LiveKit server, no network, no model download required.
+
+Run it:
+
+    uv run python examples/density_demo.py                 # N = 16
+    uv run python examples/density_demo.py --sessions 32
+    uv run python examples/density_demo.py --sessions 50 --load-vad
+
+Use --load-vad to also load the real Silero VAD in every worker (the model
+livekit-agents would load per process and OpenRTC shares). It downloads
+ONNX weights on first run, then makes the gap even wider.
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import contextlib
+import multiprocessing as mp
+import os
+import time
+
+import psutil
+
+# Stand-in for one session's live audio plus conversation state. The real
+# per-session cost is dominated by the shared-vs-per-process fixed cost, so
+# the exact buffer size is not load-bearing; it just keeps each session honest.
+_SESSION_BUFFER_MB = 5
+
+
+def _import_stack(load_vad: bool) -> None:
+    """Pay the per-process import cost that livekit-agents incurs per session."""
+    import livekit.agents  # noqa: F401  (the real wheel, ~150 MB resident)
+
+    import openrtc  # noqa: F401
+
+    if load_vad:
+        # The shared model OpenRTC loads once in prewarm and livekit-agents
+        # loads in every worker process. Widens the gap; needs a one-time
+        # weights download.
+        from livekit.plugins import silero
+
+        silero.VAD.load()
+
+
+def _process_worker(ready: object, stop: object, load_vad: bool) -> None:
+    """One subprocess == one session, the livekit-agents process-per-job model."""
+    _import_stack(load_vad)
+    _buffer = bytearray(_SESSION_BUFFER_MB * 1024 * 1024)  # noqa: F841
+    ready.set()  # type: ignore[attr-defined]
+    stop.wait()  # type: ignore[attr-defined]  hold the buffer until measured
+
+
+def measure_process_model(sessions: int, load_vad: bool) -> float:
+    """Sum resident memory of N independent worker processes (MB)."""
+    # "spawn" matches LiveKit's default executor on macOS, so each child pays
+    # the full fresh-interpreter import cost, exactly as in production.
+    ctx = mp.get_context("spawn")
+    ready_events = [ctx.Event() for _ in range(sessions)]
+    stop_event = ctx.Event()
+    procs = [
+        ctx.Process(
+            target=_process_worker, args=(ready_events[i], stop_event, load_vad)
+        )
+        for i in range(sessions)
+    ]
+    for p in procs:
+        p.start()
+    for ev in ready_events:
+        ev.wait(timeout=120)  # every worker finished importing + allocated
+
+    time.sleep(0.5)  # let resident memory settle
+    total_bytes = 0
+    for p in procs:
+        with contextlib.suppress(
+            psutil.NoSuchProcess
+        ):  # a worker may have exited early
+            total_bytes += psutil.Process(p.pid).memory_info().rss
+
+    stop_event.set()
+    for p in procs:
+        p.join()
+    return total_bytes / (1024 * 1024)
+
+
+async def measure_coroutine_model(sessions: int, load_vad: bool) -> float:
+    """Resident memory of ONE process hosting N asyncio sessions (MB)."""
+    _import_stack(load_vad)  # paid once, in this process
+
+    async def _session() -> None:
+        _buffer = bytearray(_SESSION_BUFFER_MB * 1024 * 1024)
+        try:
+            await asyncio.sleep(3600)  # stay alive until measured
+        finally:
+            del _buffer
+
+    tasks = [asyncio.create_task(_session()) for _ in range(sessions)]
+    await asyncio.sleep(0.5)  # let all sessions allocate + settle
+    rss_mb = psutil.Process(os.getpid()).memory_info().rss / (1024 * 1024)
+
+    for t in tasks:
+        t.cancel()
+    await asyncio.gather(*tasks, return_exceptions=True)
+    return rss_mb
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__.split("\n", 1)[0])
+    parser.add_argument(
+        "--sessions", type=int, default=16, help="concurrent sessions (default 16)"
+    )
+    parser.add_argument(
+        "--load-vad",
+        action="store_true",
+        help="also load real Silero VAD in every worker",
+    )
+    args = parser.parse_args()
+    n = args.sessions
+
+    print(f"\nHosting {n} concurrent voice sessions. Measuring resident memory.\n")
+
+    # Process model first so this parent process stays light; the coroutine
+    # measurement then imports the stack into this same process on purpose.
+    process_mb = measure_process_model(n, args.load_vad)
+    coroutine_mb = asyncio.run(measure_coroutine_model(n, args.load_vad))
+
+    ratio = process_mb / coroutine_mb if coroutine_mb else float("inf")
+    print(
+        f"  livekit-agents (process per session): {process_mb:8.0f} MB total   "
+        f"({process_mb / n:6.1f} MB/session)"
+    )
+    print(
+        f"  OpenRTC coroutine pool (one process): {coroutine_mb:8.0f} MB total   "
+        f"({coroutine_mb / n:6.1f} MB/session)"
+    )
+    print(f"\n  OpenRTC uses {ratio:.1f}x less memory for the same {n} sessions.\n")
+    print("  Same agent code, both ways. In OpenRTC you flip one argument:")
+    print('    AgentPool(isolation="process")    # the left column above')
+    print('    AgentPool(isolation="coroutine")  # the right column (default)\n')
+
+
+if __name__ == "__main__":
+    main()

{openrtc-0.1.0 → openrtc-0.2.2}/pyproject.toml

@@ -22,6 +22,7 @@ license = "MIT"
 requires-python = ">=3.11,<3.14"
 dependencies = [
     "livekit-agents[openai,silero,turn-detector]~=1.5",
+    "watchfiles>=0.21,<2",
 ]
 
 [project.optional-dependencies]

{openrtc-0.1.0 → openrtc-0.2.2}/src/openrtc/__init__.py

@@ -4,6 +4,7 @@ from importlib.metadata import PackageNotFoundError, version
 
 from .core.config import AgentConfig, AgentDiscoveryConfig, agent_config
 from .core.pool import AgentPool
+from .execution.file_watcher import FileChange, FileWatcher
 from .types import ProviderValue
 
 try:
@@ -18,6 +19,8 @@ __all__ = [
     "AgentConfig",
     "AgentDiscoveryConfig",
     "AgentPool",
+    "FileChange",
+    "FileWatcher",
     "ProviderValue",
     "__version__",
     "agent_config",

{openrtc-0.1.0 → openrtc-0.2.2}/src/openrtc/execution/coroutine.py

@@ -14,6 +14,8 @@ Contracts derived from:
 from __future__ import annotations
 
 import asyncio
+import contextlib
+import contextvars
 import inspect
 import logging
 import uuid
@@ -25,7 +27,7 @@ from livekit import rtc
 from livekit.agents import JobContext, JobExecutorType, JobProcess, utils
 from livekit.agents.ipc import inference_executor as inference_executor_mod
 from livekit.agents.ipc.job_executor import JobStatus
-from livekit.agents.job import RunningJobInfo
+from livekit.agents.job import RunningJobInfo, _JobContextVar
 
 if TYPE_CHECKING:
     from livekit.agents.ipc.job_executor import JobExecutor
@@ -114,6 +116,7 @@ class CoroutineJobExecutor:
         self._session_end_fnc = session_end_fnc
         self._context_factory = context_factory
         self._loop = loop
+        self._shutdown_fut: asyncio.Future[str] | None = None
 
     @property
     def id(self) -> str:
@@ -271,22 +274,79 @@ class CoroutineJobExecutor:
         self._task = loop.create_task(self._run_entrypoint(ctx))
 
     async def _run_entrypoint(self, ctx: JobContext) -> None:
+        """Run the session lifecycle, mirroring upstream ``_run_job_task``.
+
+        Establishes the job context (so ``get_job_context()`` resolves inside
+        the entrypoint and the session), holds the session open until shutdown
+        is requested (room disconnect, ``ctx.shutdown()``, or an entrypoint
+        crash), then runs the teardown sequence. Every ``JobContext`` hook is
+        treated as optional so the executor still runs with the bare stub
+        contexts that unit tests and the density benchmark pass directly.
+        """
         assert self._entrypoint_fnc is not None  # checked in launch_job
+        loop = asyncio.get_running_loop()
+        shutdown_fut: asyncio.Future[str] = loop.create_future()
+        self._shutdown_fut = shutdown_fut
+
+        def _request_shutdown(reason: str = "shutdown") -> None:
+            if not shutdown_fut.done():
+                shutdown_fut.set_result(reason)
+
+        # Per-job log fields, then the contextvar (the MAH-158 fix).
+        _on_setup = getattr(ctx, "_on_setup", None)
+        if callable(_on_setup):
+            _on_setup()
+        token: contextvars.Token[JobContext] | None = None
+        with contextlib.suppress(Exception):
+            token = _JobContextVar.set(ctx)
+
+        # Shutdown triggers (all optional for stub contexts): ctx.shutdown()
+        # via on_shutdown, and the room "disconnected" event (mirrors
+        # job_proc_lazy_main's room-disconnected handler).
+        if hasattr(ctx, "_on_shutdown"):
+
+            def _on_shutdown(reason: str = "") -> None:
+                _request_shutdown(reason or "shutdown")
+
+            ctx._on_shutdown = _on_shutdown
+        _room_on = getattr(getattr(ctx, "room", None), "on", None)
+        if callable(_room_on):
+            _room_on("disconnected", lambda *_a: _request_shutdown("room disconnected"))
+
         try:
-            await self._entrypoint_fnc(ctx)
+            try:
+                await self._entrypoint_fnc(ctx)
+            except asyncio.CancelledError:
+                if self._status is JobStatus.RUNNING:
+                    self._status = JobStatus.FAILED
+                raise
+            except Exception:
+                if self._status is JobStatus.RUNNING:
+                    self._status = JobStatus.FAILED
+                logger.exception(
+                    "entrypoint raised in CoroutineJobExecutor",
+                    extra=self.logging_extra(),
+                )
+                return
+            # Entrypoint returned cleanly. Hold a real job open until the call
+            # ends (the MAH-160 fix), then run teardown. A setup-only entrypoint
+            # (no live session) or a fake job (simulate_job, which has no live
+            # room to disconnect) completes on return instead.
+            _is_fake = getattr(ctx, "is_fake_job", None)
+            fake_job = bool(_is_fake()) if callable(_is_fake) else False
+            if (
+                getattr(ctx, "_primary_agent_session", None) is not None
+                and not fake_job
+            ):
+                try:
+                    await shutdown_fut
+                except asyncio.CancelledError:
+                    if self._status is JobStatus.RUNNING:
+                        self._status = JobStatus.FAILED
+                    raise
+                await self._teardown(ctx, shutdown_fut.result())
             if self._status is JobStatus.RUNNING:
                 self._status = JobStatus.SUCCESS
-        except asyncio.CancelledError:
-            if self._status is JobStatus.RUNNING:
-                self._status = JobStatus.FAILED
-            raise
-        except Exception:
-            if self._status is JobStatus.RUNNING:
-                self._status = JobStatus.FAILED
-            logger.exception(
-                "entrypoint raised in CoroutineJobExecutor",
-                extra=self.logging_extra(),
-            )
         finally:
             if self._session_end_fnc is not None:
                 try:
@@ -296,6 +356,43 @@ class CoroutineJobExecutor:
                         "session_end_fnc raised in CoroutineJobExecutor",
                         extra=self.logging_extra(),
                     )
+            if token is not None:
+                with contextlib.suppress(Exception):
+                    _JobContextVar.reset(token)
+
+    async def _teardown(self, ctx: JobContext, reason: str) -> None:
+        """Run the post-shutdown lifecycle (mirrors upstream ``_run_job_task``).
+
+        Closes the primary ``AgentSession``, runs ``_on_session_end`` and the
+        registered shutdown callbacks, cancels pending tasks, and cleans up.
+        Every hook is optional so stub contexts in tests and benchmarks are
+        tolerated.
+        """
+        primary = getattr(ctx, "_primary_agent_session", None)
+        if primary is not None and hasattr(primary, "aclose"):
+            with contextlib.suppress(Exception):
+                await primary.aclose()
+        _on_session_end = getattr(ctx, "_on_session_end", None)
+        if callable(_on_session_end):
+            with contextlib.suppress(Exception):
+                await _on_session_end()
+        for callback in list(getattr(ctx, "_shutdown_callbacks", None) or []):
+            try:
+                await callback(reason)
+            except Exception:
+                logger.exception(
+                    "shutdown callback raised in CoroutineJobExecutor",
+                    extra=self.logging_extra(),
+                )
+        pending = list(getattr(ctx, "_pending_tasks", None) or [])
+        if pending:
+            for task in pending:
+                task.cancel()
+            await asyncio.gather(*pending, return_exceptions=True)
+        _on_cleanup = getattr(ctx, "_on_cleanup", None)
+        if callable(_on_cleanup):
+            with contextlib.suppress(Exception):
+                _on_cleanup()
 
     def logging_extra(self) -> dict[str, Any]:
         return {"executor_id": self._id}