Kilo Code 7.8.0 detects a dead provider connection after about 10 seconds of silence.

Requests stalled on a dead connection get flagged instead of hanging forever, and the probe hits the provider's own configured endpoint over TCP so a slow local model isn't mistaken for an outage. (release) It also blocks variable references in project MCP headers nested under mcp.servers and adds a stop button to session cleanup. Distinguishing "slow" from "gone" is the part most timeout logic gets wrong.