Daemon + thin client architecture

[Aaronontheweb]

Summary

Splits Netclaw.App into two binaries and wires them together:

• netclawd (Netclaw.Daemon) — always-on daemon hosting the Akka actor system, LLM sessions, tool execution, and persistence. Exposes a functional SignalR hub at /hub/session for remote session access.
• netclaw (Netclaw.Cli) — thin CLI client with daemon management (start/stop/status/install/uninstall) and transitional in-process chat mode.

Key changes

• Project split: Netclaw.App → Netclaw.Daemon + Netclaw.Cli with shared Netclaw.Configuration
• SessionHub wired to SessionPipeline: SessionRegistry (singleton) bridges SignalR hub (transient) to Akka.Streams via IHubContext<SessionHub, ISessionHubClient> — creates sessions, accepts messages, streams output
• Daemon management: DaemonManager handles process lifecycle — binary discovery, PID file tracking, graceful SIGTERM shutdown, systemd user service registration
• Review hardening: port conflict fix (CLI uses random port), process name validation before SIGTERM, kill() return checking, output sink error logging, OfferAsync result checking, duplicate session guard
• README: updated with architecture overview and CLI reference

Integration tested

netclaw daemon start          → Daemon started (PID 1229667).
netclaw daemon status         → Daemon running (PID 1229667, uptime 0m 10s).
curl health endpoint          → "healthy"
curl SignalR negotiate        → connectionId + available transports
netclaw daemon start (again)  → Daemon already running (PID 1229667).
netclaw daemon stop           → Daemon stopped (was PID 1229667).
netclaw daemon status         → Daemon is not running.

Test plan

• dotnet build Netclaw.slnx — 0 errors, 0 warnings
• dotnet test Netclaw.slnx — 110 passed
• dotnet slopwatch analyze — 0 issues
• Daemon lifecycle: start → status → health → negotiate → duplicate guard → stop → status
• Windows cross-platform daemon management (tracked in Daemon management: Windows cross-platform support #21)

[Aaronontheweb]

Superseded by corrected review comment below. Please ignore this malformed one.

[Aaronontheweb]

Corrected review notes (prior comment had shell-escaping artifacts):

1. netclaw daemon stop can target unrelated dotnet processes because PID validation currently accepts any process named dotnet without verifying command-line identity.
2. SessionHub.SendMessage does not enforce connection ownership of sessionId, so a client that learns another session ID can inject messages.
3. Session disconnect currently disposes sessions immediately, which conflicts with the reconnect model in SPEC-011 (daemon keeps session alive for reconnection).
4. PID ownership is launcher-side only; with systemd Restart=always, PID can drift after daemon restarts and status/stop become unreliable.
5. Daemon subcommands return error text but still exit code 0, which makes automation/scripts unreliable.

Proposed fixes:

• Harden daemon PID validation (including command-line identity checks for dotnet-hosted daemon) and validate process identity in status.
• Enforce per-connection session ownership in hub/registry and add explicit session re-attach flow for reconnects.
• Move PID file authority into the daemon runtime via hosted service so restarts keep PID accurate.
• Return non-zero exit codes for daemon command failures.