One managed service

The gateway can run its inference server, SSH tunnels, native relay, and desktop login setup under one user LaunchAgent. Provider CLIs and chat executors remain separate child processes. The GitHub Actions runner and other applications are outside this service.

The same ownership rule includes the Claude Desktop integration. Its separate configuration and loopback listeners do not justify separate startup jobs. Legacy com.edihasaj.claude-* adapters, private inference and SSH forwards must be counted in a gateway audit, even though their labels do not say gateway. Node executables and SSH processes remain children of the one signed app-owned root, not additional LaunchAgents. See Claude Desktop for the read-only preview and next-start migration. Never disable a generic node Login Items row without identifying its registrations: it can contain the live Claude bridge or unrelated applications.

Claude consolidation has a stricter admission rule than a socket-idle restart: the pinned private runtime has no atomic request-drain protocol. A normal staged receipt adopts only when every legacy Claude job is unloaded, including SSH tunnels. It does not substitute an idle connection snapshot for that guarantee. An operator who has approved a maintenance interruption can stage exactly one receipt with gateway service claude-desktop --on-restart --force; that receipt may retire only its still-matching owned jobs and can interrupt Claude work.

Once any Claude component is configured, automatic primary-service restarts and upgrade activation remain pending even when no socket is currently busy. The pinned runtimes have no atomic admission gate, so a new request could otherwise arrive after an idle check. An explicit operator restart remains available for an authorized maintenance window: gateway service restart --force (see below); a plain gateway service restart refuses and names that reason. Adopted Claude SSH forwards retain their processes across wake; their existing keepalives and confirmed-exit retries recover connections instead of a forced wake restart or broad remote cleanup. Restaging without --force removes that authorization. This never forces busy Codex work to restart.

The TCP-only Claude activity probe tolerates one specific unrelated diagnostic: lsof cannot stat a CoreDevice devicefs mount under /Users/<account>/Library/Developer/CoreDevice/DeviceFS and falls back to the mount table's device number. It requires that complete three-line diagnostic; partial, changed, or additional diagnostics still refuse the probe. Darwin lsof reads mount metadata separately from its libproc socket inspection, so this filesystem warning does not hide established TCP connections. Socket output, timeouts, unexpected exits, and unknown or reused process identities remain refusals. This exception does not enable automatic restarts with Claude components configured.

Choose a task

Task	Guide
Background-item ownership and consolidation	Service ownership
Shared chat executors and native-home ownership	Shared execution
Updates, retained components and release skew	Release adoption