switch-connector: agents launched via npx never receive pushed events — hook resolves the channel port from the wrong pid

Summary

An agent connects, appears live in the room, and can read_context — but never receives an addressed message. There is no error anywhere. Seen on switch-connector 0.9.10.

Impact

Indistinguishable from a healthy but slow agent. Every observable signal says the agent is fine: the process is up, the MCP server is registered, the package resolves, the room shows the agent present. Only delivery is missing.

This is reproducible from a clean install — the npx line below is what the plugin ships, so anyone running the stock connector on a remote host hits it. It is not a one-off environment problem.

Root cause

hooks/switch_hook.py:333 resolves the channel port from the hook's own parent pid:

return os.path.join(
    os.path.expanduser("~"), ".switch", "sessions", str(os.getppid()), "port"
)

That is correct only if the runtime is a direct child of Claude Code. The comment above it says so, and even warns about the mirror case (a hook wrapped in a shell). But .mcp.json at HEAD wraps the other side:

{ "command": "npx", "args": ["-y", "@sandboxaq/switch-agent-runtime@0.3.3"] }

npx inserts npm exec → sh -c → node, so the runtime registers its port under a different pid than the hook computes. Both sides resolve the path independently — to two different paths.

Why it is silent

_notify_channel swallows the miss:

except (FileNotFoundError, ValueError):
    return

No log, no warning. This is precisely the failure CLAUDE.md's "Fail Loud, Never Fake" section exists to prevent — item 4, silently degrading to look fine.

Suggested fix

1. Don't infer the pid. Have the runtime pass its port (or the port-file path) to the hook explicitly — via env at session launch, or a fixed per-session path both sides derive from the session id rather than from a process relationship. The pid coupling is fragile regardless of npx: any future wrapper reintroduces it.
2. Warn once when the port file is absent. A single stderr line naming the path it looked for would turn days of debugging into minutes.

Workaround (confirmed working)

Point the MCP command straight at the installed binary, no npx:

{ "mcpServers": { "switch": { "command": "<prefix>/bin/switch-agent-runtime", "args": [] } } }

Delivery works immediately. Note this is overwritten by any plugin update, and the agent goes silently deaf again when it does.

Read hooks/switch_hook.py around line 333, then inspect .mcp.json and the switch-agent-runtime launch path. Reproduce the missing delivery from a clean install using the npx command and trace which session paths each process resolves. Done means the hook and runtime share a stable session-specific port path, and a missing port file emits a warning instead of failing silently.