chore: import upstream snapshot with attribution

agentlightning/execution/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from .base import ExecutionStrategy
+from .client_server import ClientServerExecutionStrategy
+from .events import ExecutionEvent, MultiprocessingEvent, ThreadingEvent
+from .shared_memory import SharedMemoryExecutionStrategy
+
+__all__ = [
+    "ExecutionStrategy",
+    "ClientServerExecutionStrategy",
+    "ExecutionEvent",
+    "ThreadingEvent",
+    "MultiprocessingEvent",
+    "SharedMemoryExecutionStrategy",
+]

agentlightning/execution/base.py

@@ -0,0 +1,64 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import logging
+from typing import Protocol
+
+from agentlightning.store.base import LightningStore
+
+from .events import ExecutionEvent
+
+logger = logging.getLogger(__name__)
+
+
+class AlgorithmBundle(Protocol):
+    """Callable bundle produced by [`Trainer`][agentlightning.Trainer].
+
+    Execution strategies treat the returned coroutine as opaque, only providing
+    the shared store instance and cooperative stop event. Bundles typically
+    encapsulate algorithm setup plus adapter and LLM proxy, etc.
+    """
+
+    async def __call__(self, store: LightningStore, event: ExecutionEvent) -> None:
+        """Execute algorithm logic using ``store`` until completion or stop."""
+
+
+class RunnerBundle(Protocol):
+    """Callable bundle wrapping runner setup and the worker loop, as opposed to the
+    [`AlgorithmBundle`][agentlightning.AlgorithmBundle]."""
+
+    async def __call__(self, store: LightningStore, worker_id: int, event: ExecutionEvent) -> None:
+        """Execute runner logic for ``worker_id`` using ``store`` and ``event``."""
+
+
+class ExecutionStrategy:
+    """Coordinate algorithm and runner bundles within a single process abstraction.
+
+    Strategies decide how many worker bundles to launch, whether to communicate
+    through shared memory or an HTTP boundary, and how to react to shutdown
+    signals. They intentionally avoid inspecting the bundle internals; instead,
+    each bundle remains responsible for its own scheduling semantics.
+
+    !!! note
+        Implementations must honor the [execute()][agentlightning.ExecutionStrategy.execute]
+        contract by propagating `KeyboardInterrupt` and ensuring resources are
+        released when an error occurs on either side of the algorithm/runner
+        pair.
+    """
+
+    def execute(self, algorithm: AlgorithmBundle, runner: RunnerBundle, store: LightningStore) -> None:
+        """Run the provided bundles using the configured orchestration model.
+
+        Args:
+            algorithm: Callable bundle responsible for algorithm execution.
+            runner: Callable bundle for runner workers.
+            store: Concrete [`LightningStore`][agentlightning.LightningStore]
+                shared across bundles.
+
+        Raises:
+            NotImplementedError: Subclasses must provide the orchestration
+                implementation.
+        """
+
+        raise NotImplementedError()

agentlightning/execution/client_server.py

@@ -0,0 +1,443 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+import asyncio
+import logging
+import multiprocessing
+import os
+import signal
+import time
+from multiprocessing.context import BaseContext
+from typing import Callable, Iterable, Literal, cast
+
+from agentlightning.env_var import LightningEnvVar, resolve_bool_env_var, resolve_int_env_var, resolve_str_env_var
+from agentlightning.store.base import LightningStore
+from agentlightning.store.client_server import LightningStoreClient, LightningStoreServer
+
+from .base import AlgorithmBundle, ExecutionStrategy, RunnerBundle
+from .events import ExecutionEvent, MultiprocessingEvent
+
+logger = logging.getLogger(__name__)
+
+
+class ClientServerExecutionStrategy(ExecutionStrategy):
+    """Run algorithm and runner bundles as separate processes over HTTP.
+
+    Execution Roles:
+
+    - `"algorithm"`: Start [`LightningStoreServer`][agentlightning.LightningStoreServer]
+      in-process and execute the algorithm bundle against it.
+    - `"runner"`: Connect to an existing server with
+      [`LightningStoreClient`][agentlightning.LightningStoreClient] and run the
+      runner bundle locally (spawning multiple processes when requested).
+    - `"both"`: Spawn runner processes first, then execute the algorithm and
+      server on the same machine. This mode orchestrates the full loop locally.
+
+    When `role == "both"` you may choose which side runs on the main process
+    via `main_process`. The runner-on-main option is limited to
+    `n_runners == 1` because each additional runner requires its own event
+    loop and process.
+
+    !!! warning
+        When `main_process == "runner"` the algorithm and HTTP server execute
+        in a child process. Store mutations remain isolated inside that process,
+        so the original store instance passed to
+        [execute()][agentlightning.ExecutionStrategy.execute] is not updated.
+
+    Abort Model (four-step escalation):
+
+    1. Cooperative stop. Every bundle receives a shared
+       [`MultiprocessingEvent`][agentlightning.MultiprocessingEvent] (`stop_evt`).
+       Any failure flips the event so peers can exit cleanly. Ctrl+C on the main
+       process also sets the flag.
+    2. KeyboardInterrupt synthesis. Remaining subprocesses receive ``SIGINT`` to
+       trigger `KeyboardInterrupt` handlers.
+    3. Termination. Stubborn processes are asked to ``terminate()``
+       (`SIGTERM` on POSIX).
+    4. Kill. As a last resort `kill()` is invoked (`SIGKILL` on POSIX).
+
+    This mirrors the semantics implemented in
+    [`SharedMemoryExecutionStrategy`][agentlightning.SharedMemoryExecutionStrategy]
+    but adapts them to multiple processes and the HTTP client/server boundary.
+    """
+
+    alias: str = "cs"
+
+    def __init__(
+        self,
+        role: Literal["algorithm", "runner", "both"] | None = None,
+        server_host: str | None = None,
+        server_port: int | None = None,
+        n_runners: int = 1,
+        graceful_timeout: float = 10.0,
+        terminate_timeout: float = 10.0,
+        main_process: Literal["algorithm", "runner"] = "algorithm",
+        managed_store: bool | None = None,
+        allowed_exit_codes: Iterable[int] = (0, -15),
+    ) -> None:
+        """Configure the strategy.
+
+        Args:
+            role: Which side(s) to run in this process. When omitted, the
+                `AGL_CURRENT_ROLE` environment variable is used.
+            server_host: Interface the HTTP server binds to when running the
+                algorithm bundle locally. Defaults to `AGL_SERVER_HOST`
+                or `"localhost"` if unset.
+            server_port: Port for the HTTP server in "algorithm"/"both" modes.
+                Defaults to `AGL_SERVER_PORT` or `4747` if unset.
+            n_runners: Number of runner processes to spawn in "runner"/"both".
+            graceful_timeout: How long to wait (seconds) after setting the stop
+                event before escalating to signals.
+            terminate_timeout: How long to wait between escalation steps beyond
+                the cooperative phase (re-used for SIGINT, terminate, and kill).
+            main_process: Which bundle runs on the main process when
+                `role == "both"`. `"runner"` requires `n_runners == 1` and is
+                primarily intended for debugging.
+            managed_store: When `True` (default) the strategy constructs
+                LightningStore client/server wrappers automatically. When
+                `False` the provided `store` is passed directly to the
+                bundles, allowing callers to manage store wrappers manually.
+            allowed_exit_codes: Allowed exit codes for subprocesses.
+                By default, runner can exit gracefully with code 0 or terminated
+                by SIGTERM (-15).
+        """
+        resolved_role = resolve_str_env_var(LightningEnvVar.AGL_CURRENT_ROLE, override=role, fallback="both")
+        if resolved_role not in ("algorithm", "runner", "both"):
+            raise ValueError("role must be one of 'algorithm', 'runner', or 'both'")
+        self.role: Literal["algorithm", "runner", "both"] = resolved_role
+        self.n_runners = n_runners
+        self.server_host = resolve_str_env_var(
+            LightningEnvVar.AGL_SERVER_HOST, override=server_host, fallback="localhost"
+        )
+        self.server_port = resolve_int_env_var(LightningEnvVar.AGL_SERVER_PORT, override=server_port, fallback=4747)
+        self.graceful_timeout = graceful_timeout
+        self.terminate_timeout = terminate_timeout
+        if main_process not in ("algorithm", "runner"):
+            raise ValueError("main_process must be 'algorithm' or 'runner'")
+        if main_process == "runner":
+            if self.role != "both":
+                raise ValueError("main_process='runner' is only supported when role='both'")
+            if n_runners != 1:
+                raise ValueError("main_process='runner' requires n_runners to be 1")
+        self.main_process = main_process
+        self.managed_store = resolve_bool_env_var(
+            LightningEnvVar.AGL_MANAGED_STORE, override=managed_store, fallback=True
+        )
+        self.allowed_exit_codes = tuple(allowed_exit_codes)
+
+    async def _execute_algorithm(
+        self, algorithm: AlgorithmBundle, store: LightningStore, stop_evt: ExecutionEvent
+    ) -> None:
+        wrapper_store: LightningStore | None = None
+        if self.managed_store:
+            logger.info("Starting LightningStore server on %s:%s", self.server_host, self.server_port)
+            wrapper_store = LightningStoreServer(store, host=self.server_host, port=self.server_port)
+            server_started = False
+        else:
+            wrapper_store = store
+            server_started = False
+
+        try:
+            if self.managed_store and isinstance(wrapper_store, LightningStoreServer):
+                await wrapper_store.start()
+                server_started = True
+                logger.debug("Algorithm bundle starting against endpoint %s", wrapper_store.endpoint)
+            await algorithm(wrapper_store, stop_evt)
+            logger.debug("Algorithm bundle completed successfully")
+        except asyncio.CancelledError:
+            logger.info("Algorithm received CancelledError; signaling stop event")
+            stop_evt.set()
+            raise
+        except KeyboardInterrupt:
+            logger.warning("Algorithm received KeyboardInterrupt; signaling stop event")
+            stop_evt.set()
+            raise
+        except BaseException:
+            logger.exception("Algorithm bundle crashed; signaling stop event")
+            stop_evt.set()
+            raise
+        finally:
+            if self.managed_store and isinstance(wrapper_store, LightningStoreServer) and server_started:
+                try:
+                    await wrapper_store.stop()
+                except Exception:
+                    logger.exception("Error stopping LightningStore server")
+                else:
+                    logger.debug("LightningStore server shutdown completed")
+
+    async def _execute_runner(
+        self,
+        runner: RunnerBundle,
+        worker_id: int,
+        store: LightningStore,
+        stop_evt: ExecutionEvent,
+    ) -> None:
+        if self.managed_store:
+            # If managed, we actually do not use the provided store
+            client_store = LightningStoreClient(f"http://{self.server_host}:{self.server_port}")
+        else:
+            client_store = store
+        try:
+            if self.managed_store:
+                logger.debug("Runner %s connecting to server at %s:%s", worker_id, self.server_host, self.server_port)
+            else:
+                logger.debug("Runner %s executing with provided store", worker_id)
+            await runner(client_store, worker_id, stop_evt)
+            logger.debug("Runner %s completed successfully", worker_id)
+        except asyncio.CancelledError:
+            logger.debug("Runner %s received CancelledError; signaling stop event", worker_id)
+            stop_evt.set()
+            raise
+        except KeyboardInterrupt:
+            logger.warning("Runner %s received KeyboardInterrupt; signaling stop event", worker_id)
+            stop_evt.set()
+            raise
+        except BaseException:
+            logger.exception("Runner %s crashed; signaling stop event", worker_id)
+            stop_evt.set()
+            raise
+        finally:
+            if self.managed_store and isinstance(client_store, LightningStoreClient):
+                try:
+                    await client_store.close()
+                except Exception:
+                    logger.exception("Error closing LightningStore client for runner %s", worker_id)
+                else:
+                    logger.debug("Runner %s closed LightningStore client", worker_id)
+
+    def _spawn_runners(
+        self,
+        runner: RunnerBundle,
+        store: LightningStore,
+        stop_evt: ExecutionEvent,
+        *,
+        ctx: BaseContext,
+    ) -> list[multiprocessing.Process]:
+        """Used when `role == "runner"` or `role == "both"` and `n_runners > 1`."""
+        processes: list[multiprocessing.Process] = []
+
+        def _runner_sync(runner: RunnerBundle, worker_id: int, store: LightningStore, stop_evt: ExecutionEvent) -> None:
+            # Runners are executed in child processes; each process owns its own
+            # event loop to keep the asyncio scheduler isolated.
+            try:
+                asyncio.run(self._execute_runner(runner, worker_id, store, stop_evt))
+            except KeyboardInterrupt:
+                logger.warning("Runner (asyncio) %s received KeyboardInterrupt; exiting gracefully", worker_id)
+            except BaseException as exc:
+                logger.exception("Runner (asyncio) %s crashed by %s; signaling stop event", worker_id, exc)
+                raise
+
+        for i in range(self.n_runners):
+            process = cast(
+                multiprocessing.Process,
+                ctx.Process(target=_runner_sync, args=(runner, i, store, stop_evt), name=f"runner-{i}"),  # type: ignore
+            )
+            process.start()
+            logger.debug("Spawned runner process %s (pid=%s)", process.name, process.pid)
+            processes.append(process)
+
+        return processes
+
+    def _spawn_algorithm_process(
+        self,
+        algorithm: AlgorithmBundle,
+        store: LightningStore,
+        stop_evt: ExecutionEvent,
+        *,
+        ctx: BaseContext,
+    ) -> multiprocessing.Process:
+        """Used when `main_process == "runner"`."""
+
+        def _algorithm_sync(algorithm: AlgorithmBundle, store: LightningStore, stop_evt: ExecutionEvent) -> None:
+            try:
+                asyncio.run(self._execute_algorithm(algorithm, store, stop_evt))
+            except KeyboardInterrupt:
+                logger.warning("Algorithm (asyncio.run) received KeyboardInterrupt; exiting gracefully")
+            except BaseException as exc:
+                logger.exception("Algorithm (asyncio.run) crashed by %s; signaling stop event", exc)
+                raise
+
+        process = cast(
+            multiprocessing.Process,
+            ctx.Process(target=_algorithm_sync, args=(algorithm, store, stop_evt), name="algorithm"),  # type: ignore
+        )
+        process.start()
+        logger.debug("Spawned algorithm process %s (pid=%s)", process.name, process.pid)
+        return process
+
+    def _join_until_deadline(
+        self,
+        processes: Iterable[multiprocessing.Process],
+        timeout: float,
+    ) -> list[multiprocessing.Process]:
+        """Join ``processes`` until ``timeout`` elapses, returning those still alive."""
+        deadline = time.monotonic() + timeout
+        still_alive: list[multiprocessing.Process] = []
+        for process in processes:
+            remaining = deadline - time.monotonic()
+            if remaining > 0:
+                process.join(remaining)
+            else:
+                process.join(0)
+            if process.is_alive():
+                still_alive.append(process)
+        return still_alive
+
+    def _signal_processes(
+        self,
+        processes: Iterable[multiprocessing.Process],
+        action: Callable[[multiprocessing.Process], None],
+    ) -> None:
+        """Invoke ``action`` on each process while suppressing individual failures."""
+        for process in processes:
+            try:
+                action(process)
+            except Exception:
+                logger.exception("Error signaling process %s (pid=%s)", process.name, process.pid)
+
+    def _shutdown_processes(
+        self,
+        processes: list[multiprocessing.Process],
+        stop_evt: ExecutionEvent,
+    ) -> None:
+        """4-step escalation shutdown of ``processes``."""
+        if not processes:
+            logger.debug("No subprocesses to shutdown")
+            return
+
+        if not stop_evt.is_set():
+            logger.debug("Sending cooperative stop signal to subprocesses")
+            stop_evt.set()
+        else:
+            logger.debug("Stop event already set; waiting for subprocesses to exit")
+
+        alive = self._join_until_deadline(processes, self.graceful_timeout)
+        if not alive:
+            return
+
+        logger.warning(
+            "Subprocesses still alive after cooperative wait; sending SIGINT to %s",
+            ", ".join(p.name or str(p.pid) for p in alive),
+        )
+        # SIGINT is not reliable on Windows, but we do not consider such case yet.
+        self._signal_processes(alive, lambda p: os.kill(cast(int, p.pid), signal.SIGINT))
+        alive = self._join_until_deadline(alive, self.terminate_timeout)
+        if not alive:
+            return
+
+        logger.warning(
+            "Subprocesses still alive after SIGINT wait; sending terminate() to %s",
+            ", ".join(p.name or str(p.pid) for p in alive),
+        )
+        self._signal_processes(alive, lambda p: p.terminate())
+
+        alive = self._join_until_deadline(alive, self.terminate_timeout)
+        if not alive:
+            return
+
+        logger.error(
+            "Subprocesses still alive after terminate(); sending kill() to %s",
+            ", ".join(p.name or str(p.pid) for p in alive),
+        )
+        self._signal_processes(alive, lambda p: p.kill())
+        alive = self._join_until_deadline(alive, self.terminate_timeout)
+
+        if alive:
+            logger.error(
+                "Subprocesses failed to exit even after kill(): %s", ", ".join(p.name or str(p.pid) for p in alive)
+            )
+
+    def _check_process_exitcodes(self, processes: Iterable[multiprocessing.Process]) -> None:
+        """Raise an error if any managed process exited with a non-zero status."""
+        failed = [p for p in processes if p.exitcode not in self.allowed_exit_codes + (None,)]
+        if failed:
+            formatted = ", ".join(f"{p.name or p.pid} (exitcode={p.exitcode})" for p in failed)
+            raise RuntimeError(f"Subprocesses failed with unexpected exit codes: {formatted}")
+
+    def execute(self, algorithm: AlgorithmBundle, runner: RunnerBundle, store: LightningStore) -> None:
+        logger.info(
+            "Starting client-server execution with %d runner(s) [role=%s, main_process=%s]",
+            self.n_runners,
+            self.role,
+            self.main_process,
+        )
+
+        # Re-use the active multiprocessing context so the event and processes
+        # agree on the start method (fork/spawn/forkserver).
+        ctx = multiprocessing.get_context()
+        stop_evt = MultiprocessingEvent(ctx=ctx)
+        # Track spawned processes so we can enforce termination ordering and
+        # surface non-zero exit codes back to the caller.
+        processes: list[multiprocessing.Process] = []
+
+        exception: BaseException | None = None
+        keyboard_interrupt = False
+
+        try:
+            if self.role == "algorithm":
+                logger.info("Running algorithm solely...")
+                asyncio.run(self._execute_algorithm(algorithm, store, stop_evt))
+            elif self.role == "runner":
+                if self.n_runners == 1:
+                    logger.info("Running runner solely...")
+                    asyncio.run(self._execute_runner(runner, 0, store, stop_evt))
+                else:
+                    logger.info("Spawning runner processes...")
+                    processes = self._spawn_runners(runner, store, stop_evt, ctx=ctx)
+                    # Wait for the processes to finish naturally.
+                    for process in processes:
+                        process.join()
+                    self._check_process_exitcodes(processes)
+            elif self.role == "both":
+                if self.main_process == "algorithm":
+                    logger.info("Spawning runner processes...")
+                    processes = self._spawn_runners(runner, store, stop_evt, ctx=ctx)
+                    try:
+                        logger.info("Running algorithm...")
+                        asyncio.run(self._execute_algorithm(algorithm, store, stop_evt))
+                    finally:
+                        # Always request the runner side to unwind once the
+                        # algorithm/server portion finishes (successfully or not).
+                        stop_evt.set()
+                else:  # main_process == "runner"
+                    if self.n_runners > 1:
+                        raise ValueError("main_process='runner' requires n_runners to be 1")
+
+                    logger.info("Spawning algorithm process...")
+                    algorithm_process = self._spawn_algorithm_process(algorithm, store, stop_evt, ctx=ctx)
+                    processes = [algorithm_process]
+
+                    # Run the lone runner cooperatively in-process so users can
+                    # attach a debugger. The algorithm + HTTP server live in
+                    # the background process spawned above (the provided
+                    # store must therefore be picklable when using spawn).
+                    logger.info("Running runner...")
+                    asyncio.run(self._execute_runner(runner, 0, store, stop_evt))
+
+                    # Wait for the algorithm process to finish.
+                    algorithm_process.join()
+            else:
+                raise ValueError(f"Unknown role: {self.role}")
+        except KeyboardInterrupt:
+            logger.warning("KeyboardInterrupt received; initiating shutdown")
+            stop_evt.set()
+            keyboard_interrupt = True
+        except BaseException as exc:
+            logger.exception("Unhandled exception in execute method")
+            stop_evt.set()
+            # Preserve the original exception so we can avoid masking it during
+            # the cleanup phase.
+            exception = exc
+            raise
+        finally:
+            logger.info("Shutting down subprocesses")
+            self._shutdown_processes(processes, stop_evt)
+            if processes:
+                try:
+                    self._check_process_exitcodes(processes)
+                except RuntimeError as err:
+                    if exception is not None or keyboard_interrupt:
+                        # We already propagate/handled a different failure, so
+                        # emit a warning instead of raising a secondary error.
+                        logger.warning("Subprocesses ended abnormally during shutdown: %s", err)
+                    else:
+                        raise

agentlightning/execution/events.py

@@ -0,0 +1,69 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+import multiprocessing as mp
+import threading
+from multiprocessing.context import BaseContext
+from typing import Optional, Protocol
+
+
+class ExecutionEvent(Protocol):
+    """Protocol capturing the cooperative stop contract shared by strategies.
+
+    Implementations mirror the API of ``threading.Event`` and
+    ``multiprocessing.Event`` so the rest of the execution layer can remain
+    agnostic to the underlying concurrency primitive.
+
+    Methods:
+
+        set: Signal cancellation. The call must be idempotent.
+        clear: Reset the event to the unsignaled state.
+        is_set: Return ``True`` when cancellation has been requested.
+        wait: Block until the event is signaled or an optional timeout elapses.
+    """
+
+    def set(self) -> None: ...
+    def clear(self) -> None: ...
+    def is_set(self) -> bool: ...
+    def wait(self, timeout: Optional[float] = None) -> bool: ...
+
+
+class ThreadingEvent:
+    """Thread-safe implementation of [`ExecutionEvent`][agentlightning.ExecutionEvent]."""
+
+    __slots__ = ("_evt",)
+
+    def __init__(self) -> None:
+        self._evt = threading.Event()
+
+    def set(self) -> None:
+        self._evt.set()
+
+    def clear(self) -> None:
+        self._evt.clear()
+
+    def is_set(self) -> bool:
+        return self._evt.is_set()
+
+    def wait(self, timeout: Optional[float] = None) -> bool:
+        return self._evt.wait(timeout)
+
+
+class MultiprocessingEvent:
+    """Process-safe implementation of [`ExecutionEvent`][agentlightning.ExecutionEvent]."""
+
+    __slots__ = ("_evt",)
+
+    def __init__(self, *, ctx: Optional[BaseContext] = None) -> None:
+        self._evt = (ctx or mp).Event()
+
+    def set(self) -> None:
+        self._evt.set()
+
+    def clear(self) -> None:
+        self._evt.clear()
+
+    def is_set(self) -> bool:
+        return self._evt.is_set()
+
+    def wait(self, timeout: Optional[float] = None) -> bool:
+        return self._evt.wait(timeout)

agentlightning/execution/inter_process.py

@@ -0,0 +1,16 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from .base import ExecutionStrategy
+
+
+class InterProcessExecutionStrategy(ExecutionStrategy):
+    """Placeholder strategy for future inter-process primitives.
+
+    The class exists to reserve the `ipc` alias and make the planned
+    implementation discoverable. Attempting to use it today will raise
+    `NotImplementedError` once the execution contract is finalized.
+    """
+
+    alias: str = "ipc"
+
+    # TODO: to be implemented

agentlightning/execution/shared_memory.py

@@ -0,0 +1,282 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+import asyncio
+import logging
+import threading
+from contextlib import suppress
+from queue import SimpleQueue
+from typing import Any, Awaitable, Callable, List, Literal, Optional, Tuple
+
+from agentlightning.env_var import LightningEnvVar, resolve_bool_env_var
+from agentlightning.store.base import LightningStore
+from agentlightning.store.threading import LightningStoreThreaded
+
+from .base import AlgorithmBundle, ExecutionStrategy, RunnerBundle
+from .events import ExecutionEvent, ThreadingEvent
+
+logger = logging.getLogger(__name__)
+
+
+class SharedMemoryExecutionStrategy(ExecutionStrategy):
+    """Execute bundles in a single process with cooperative worker threads.
+
+    Stop Model:
+
+    - All bundles share one [`ThreadingEvent`][agentlightning.ThreadingEvent]
+      named `stop_evt`.
+    - Only the main thread receives `KeyboardInterrupt`. When Ctrl+C occurs we
+      set `stop_evt`.
+    - Any exception raised inside a bundle sets `stop_evt` so other threads can
+      unwind cooperatively.
+    - Once the bundle running on the main thread exits successfully the
+      treatment depends on `main_thread`:
+        - `"algorithm"`: the runners are asked to stop by setting `stop_evt`.
+        - `"runner"`: the algorithm keeps running until it exits naturally.
+    - Background threads are marked as daemons. We join them briefly and log any
+      stragglers before shutting down.
+
+    !!! note
+        Signals other than `SIGINT` (such as `SIGTERM`) are not intercepted;
+        Python's default behavior for those signals is preserved.
+    """
+
+    alias: str = "shm"
+
+    def __init__(
+        self,
+        n_runners: int = 1,
+        main_thread: Literal["algorithm", "runner"] = "runner",
+        join_timeout: float = 15.0,
+        graceful_delay: float = 5.0,
+        poll_interval: float = 0.05,
+        managed_store: bool | None = None,
+    ) -> None:
+        if main_thread not in ("algorithm", "runner"):
+            raise ValueError("main_thread must be 'algorithm' or 'runner'")
+        if main_thread == "runner" and n_runners != 1:
+            raise ValueError(
+                "When main_thread is 'runner', n_runners must be 1. "
+                "Either use 'algorithm' on the main thread or set n_runners to 1."
+            )
+        self.n_runners = n_runners
+        self.main_thread = main_thread
+        self.join_timeout = join_timeout
+        self.graceful_delay = graceful_delay
+        self.poll_interval = poll_interval
+        self.managed_store = resolve_bool_env_var(
+            LightningEnvVar.AGL_MANAGED_STORE, override=managed_store, fallback=True
+        )
+
+    async def _run_until_completed_or_canceled(self, coro: Awaitable[Any], stop_evt: ExecutionEvent) -> Any:
+        """Run `coro` until it finishes or a cooperative stop is requested.
+
+        Control flow:
+
+        1. Start the bundle coroutine as `task`.
+        2. Launch a watcher that polls `stop_evt` without blocking the loop.
+        3. When the stop event flips:
+            a. Give the bundle `graceful_delay` seconds to finish on its own,
+               because well-behaved bundles will check the event and return.
+            b. Cancel the bundle task if it is still running after the grace
+               period.
+        4. Await both tasks and swallow `CancelledError` where appropriate.
+
+        This is a *backup* mechanism for bundles that might not poll the event
+        frequently; cooperative shutdown (checking `stop_evt` inside the
+        bundle) remains the preferred approach.
+        """
+        task: asyncio.Task[Any] = asyncio.create_task(coro)  # type: ignore
+        task_exception: Optional[BaseException] = None
+
+        async def watcher() -> None:
+            # Poll the threading event without blocking the event loop. Using a
+            # background thread via ``asyncio.to_thread`` makes cancellation
+            # difficult because ``ThreadingEvent.wait`` is not interruptible.
+            # Instead we cooperatively check the flag from the loop so the
+            # watcher task stays cancellable and tests don't hang when the
+            # bundle finishes naturally before the stop event is set.
+            while not stop_evt.is_set():
+                await asyncio.sleep(self.poll_interval)
+
+            # Grace period: let a cooperative bundle exit on its own.
+            try:
+                # At this point of waiting, the main task should already see the stop event.
+                await asyncio.wait_for(asyncio.shield(task), timeout=self.graceful_delay)  # type: ignore
+                logger.debug("Bundle finished by itself during grace period.")
+                return  # bundle finished by itself during grace period
+            except asyncio.TimeoutError:
+                # Still running after the grace window.
+                pass
+            except asyncio.CancelledError:
+                # If someone else canceled the task already, we're done.
+                logger.debug("Bundle already canceled by someone else; exiting watcher.")
+                return
+
+            # Still running after the grace window: cancel it.
+            if not task.done():
+                logger.debug("Graceful delay elapsed; canceling bundle task...")
+                task.cancel()
+
+        watcher_task = asyncio.create_task(watcher())
+        result: Any = None
+
+        try:
+            # We don't wait on FIRST_COMPLETED here, because we want the watcher
+            # to be able to grant a grace window after stop_evt flips.
+            await asyncio.wait(
+                {task, watcher_task}, return_when=asyncio.FIRST_COMPLETED
+            )  # pyright: ignore[reportUnknownArgumentType]
+        finally:
+            # If the main task hasn't completed yet (e.g., watcher scheduled cancel),
+            # finish the cancellation handshake.
+            if not task.done():
+                try:
+                    await asyncio.wait_for(task, timeout=self.graceful_delay)  # second chance
+                except asyncio.TimeoutError:
+                    logger.error(
+                        "Bundle task did not stop after cancellation; abandoning task."
+                        "This thread could live until the process exits."
+                    )
+                    # We return without awaiting it. asyncio.run will still try to cancel
+                    # pending tasks on loop close; if the task ignores cancellation, this
+                    # thread may still stick. It's the best we can do in Python.
+                    # We don't raise an exception here, but the thread could be a zombie.
+                    return result
+            else:
+                # Task completed naturally; retrieve result.
+                try:
+                    result = await task  # type: ignore
+                except asyncio.CancelledError:
+                    pass
+                except BaseException as exc:
+                    task_exception = exc
+
+            watcher_task.cancel()
+            with suppress(asyncio.CancelledError):
+                await watcher_task
+
+        if task_exception is not None:
+            raise task_exception
+
+        return result  # type: ignore
+
+    def _run_algorithm(
+        self,
+        algorithm: AlgorithmBundle,
+        store: LightningStore,
+        stop_evt: ExecutionEvent,
+        thread_exceptions: Optional[SimpleQueue[BaseException]],
+    ) -> None:
+        try:
+            asyncio.run(self._run_until_completed_or_canceled(algorithm(store, stop_evt), stop_evt))
+        except asyncio.CancelledError:
+            logger.info("Algorithm bundle canceled due to stop signal.")
+        except BaseException as exc:
+            logger.exception("Algorithm bundle crashed; signaling stop to others.")
+            if thread_exceptions is not None:
+                thread_exceptions.put(exc)
+            stop_evt.set()
+            raise
+
+    def _run_runner(
+        self,
+        runner: RunnerBundle,
+        store: LightningStore,
+        worker_id: int,
+        stop_evt: ExecutionEvent,
+        thread_exceptions: Optional[SimpleQueue[BaseException]],
+    ) -> None:
+        try:
+            asyncio.run(self._run_until_completed_or_canceled(runner(store, worker_id, stop_evt), stop_evt))
+        except asyncio.CancelledError:
+            logger.info("Runner bundle (worker_id=%s) canceled due to stop signal.", worker_id)
+        except BaseException as exc:
+            logger.exception("Runner bundle crashed (worker_id=%s); signaling stop to others.", worker_id)
+            if thread_exceptions is not None:
+                thread_exceptions.put(exc)
+            stop_evt.set()
+            raise
+
+    def execute(self, algorithm: AlgorithmBundle, runner: RunnerBundle, store: LightningStore) -> None:
+        logger.info(
+            "Starting shm execution with %d runner(s); main thread runs '%s'",
+            self.n_runners,
+            self.main_thread,
+        )
+
+        # Create stop event and thread-safe store.
+        stop_evt = ThreadingEvent()
+        if self.managed_store:
+            thread_safe_store = LightningStoreThreaded(store)
+        else:
+            thread_safe_store = store
+
+        thread_exceptions: SimpleQueue[BaseException] = SimpleQueue()
+        raised_from_thread: Optional[BaseException] = None
+
+        def make_thread(name: str, target: Callable[..., Any], args: Tuple[Any, ...]) -> threading.Thread:
+            t = threading.Thread(name=name, target=target, args=args, daemon=True)
+            t.start()
+            return t
+
+        threads: List[threading.Thread] = []
+
+        try:
+            if self.main_thread == "algorithm":
+                # Start runner threads; algorithm runs on main thread.
+                for i in range(self.n_runners):
+                    thread = make_thread(
+                        name=f"runner-{i}",
+                        target=self._run_runner,
+                        args=(runner, thread_safe_store, i, stop_evt, thread_exceptions),
+                    )
+                    threads.append(thread)
+
+                # Ctrl+C here raises KeyboardInterrupt on this stack.
+                # Main thread doesn't need to collect exceptions.
+                self._run_algorithm(algorithm, thread_safe_store, stop_evt, None)
+
+                # If algo finishes naturally, request runners to stop.
+                stop_evt.set()
+
+            else:  # main_thread == "runner"
+                # Start algorithm in background; runner runs on main thread.
+                thread = make_thread(
+                    name="algorithm",
+                    target=self._run_algorithm,
+                    args=(algorithm, thread_safe_store, stop_evt, thread_exceptions),
+                )
+                threads.append(thread)
+
+                # Ctrl+C here raises KeyboardInterrupt on this stack.
+                # Main thread doesn't need to collect exceptions.
+                self._run_runner(runner, thread_safe_store, 0, stop_evt, None)
+
+                # If runner finishes naturally, WAIT FOR ALGORITHM TO FINISH.
+                thread.join()
+
+            if not thread_exceptions.empty():
+                raised_from_thread = thread_exceptions.get()
+
+        except KeyboardInterrupt:
+            logger.warning("KeyboardInterrupt received on main thread; initiating cooperative shutdown...")
+            stop_evt.set()
+        finally:
+            # Attempt a clean join; if some threads don't comply, log and move on.
+            for t in threads:
+                logger.debug("Joining thread %s...", t.name)
+                t.join(timeout=self.join_timeout)
+
+            alive = [t.name for t in threads if t.is_alive()]
+            if alive:
+                logger.error(
+                    "Threads still alive after %.1fs: %s. They are daemons; continuing shutdown.",
+                    self.join_timeout,
+                    ", ".join(alive),
+                )
+
+            if raised_from_thread is None and not thread_exceptions.empty():
+                raised_from_thread = thread_exceptions.get()
+
+        if raised_from_thread is not None:
+            raise raised_from_thread