proclaunch

Start and stop long-running local background processes, detached and windowless, with process identity (pid and creation time), exclusive start reservations, readiness probes and an optional OS supervisor (Windows Task Scheduler). It defines its own BackgroundLauncher / Supervisor ports and launch types (proclaunch.domain, proclaunch.ports).

from proclaunch.api import launch_background, start_service, stop_service
from proclaunch.domain import ProcessSpec

result = launch_background(ProcessSpec(("python", "-m", "myapp")))   # a LaunchResult; never raises
if not result.ok:
    print(result.error)

A named service (ProcessSpec(name=...)) is recorded so a second start is ALREADY_RUNNING, and a stop only acts when the recorded identity still matches (a reused pid is left alone).

Tests: bind proclaunch.testing.FakeLauncher with proclaunch.api.use_launcher(...) and FakeSupervisors with use_supervisor(...).

Replacing one collaborator

The operations keep proclaunch's invariants (exclusive reservation, pid + creation-time identity, the record); what starts the process, how identity is read, where records live and the clock can each be replaced on its own:

from proclaunch.api import PtySpawner, default_manager, start_service

# a different way to start the process; reservation, identity and records stay proclaunch's
start_service(ProcessSpec(("claude", "remote-control"), name="rc"), state_dir=state, spawner=PtySpawner())

manager = default_manager(state, spawner=PtySpawner())   # or hold a manager and use it repeatedly

proclaunch.api.__all__ is the supported surface: the operations, the values they return, the ports you may implement (ProcessSpawner, Probe, Supervisor, ...) and the ready-made implementations (SubprocessSpawner, PtySpawner, HttpProbe, TcpProbe, CallableProbe). A launcher bound with use_launcher (a test seam) takes precedence over an injected spawner, so code that injects one cannot start a real process inside a test.