Task and process recovery
recovered
Section titled “recovered”The persisted process identity still matches a live process. Nerve can stop or restart it, but cannot reconnect old stdout/stderr pipes. Output captured before application restart stays frozen; only newly observable state is available.
interrupted
Section titled “interrupted”The process is verified gone. Restart creates a new immutable run in the same task definition lineage. Inspect the prior run’s terminal state and application logs before restarting.
recovery_unknown
Section titled “recovery_unknown”Nerve cannot prove that the PID still belongs to the original process. It blocks destructive PID actions to avoid signaling an unrelated reused process. Verify externally and start a new task rather than forcing through Nerve.
Readiness never arrives
Section titled “Readiness never arrives”Check the readiness URL/pattern and task logs. A server listening on a different interface/port can run correctly while missing the declared check. Use task log filters instead of polling repeatedly from the agent.
Stop does not complete
Section titled “Stop does not complete”Nerve attempts graceful process-tree termination then bounded escalation. A run stuck in stopping offers Force kill immediately, since you already asked to stop it. A verified recovered run confirms before killing, because you may not have asked to terminate it. Force kill can lose buffered output and process cleanup, but the terminal run record can then be removed or its current saved task definition can be run again.
Platform security software or child detachment can still interfere. Nerve never offers force kill for recovery_unknown; confirm identity before manual OS termination.
Quitting Nerve
Section titled “Quitting Nerve”Quitting the desktop application immediately terminates active local task process trees before the owned daemon shuts down. Closing a window to the system tray does not quit Nerve or stop tasks.
Agent run recovery is separate
Section titled “Agent run recovery is separate”Conversation runs use durable checkpoints, retry policy, and manual continuation. A background process being recovered does not automatically mean the model run is resumable, and vice versa.