[grytchat]

Part of the Gryt MCP (GRYT-2005). This adds what the MCP's server_status reads: a server:diagnostics event behind a new view_diagnostics permission.

It returns:

• who's in each voice channel, by nickname, with mute, deafen, camera, screen shares and the latency stats their client already reports
• the server version, uptime and online count
• the SFU connection state
• the SFU's own gryt_sfu_* numbers, read from SFU_METRICS_URL. Null when it's unset or the SFU doesn't answer within 2 seconds

view_diagnostics is owner-only to begin with, like manage_bots: admins don't get it, and the backfill (schema 11) gives it to roles holding manage_server. A bot gets it only when it asks and an owner leaves it ticked.

The self-hosted bundle sets SFU_METRICS_URL to the SFU's metrics listener on 127.0.0.1, and example.env documents it, commented out.

Look at

• What collectDiagnostics exposes. Nicknames only, with no user ids and no IPs. Latency is what each client already sends to the others in its room.
• SFU_METRICS_URL is fetched server-side. It comes from the operator's env, not from a request, so it isn't an SSRF surface.
• No rate limit on the event, since only owner-level callers reach it. Say if you want one anyway.

Tested

• the gate test (refuses without view_diagnostics, accepts with it), backfill and migration tests, and a parser test for the metrics text
• end to end with a real SFU and the MCP from bot#19: server_status returned live gryt_sfu_* numbers. Nobody was in voice, so the room list was empty.

Client label: client#1141. Docs: docs#181. Should go out on beta first.

🤖 Generated with Claude Code