Summary

Adds a configurable channelHealthRestartMode option to the channel health-monitor that allows lightweight channel reconnects instead of full stop/start cycles when a stale socket is detected.

Closes openclaw#57569

Problem

When a channel (Discord, Telegram, etc.) reports a stale-socket, the health-monitor performs a stopChannel → startChannel cycle. This is correct for hard failures, but for transient stale-socket conditions it has a costly side effect: Anthropic's prompt cache gets cold-started, forcing a full context re-send (~120k tokens) on the next turn.

Real-world impact: ~240× token usage spike after any stale-socket event, until the cache warms back up over 2-3 turns.

Solution

Add a restartMode that controls how the health-monitor handles channel recovery:

// Global default — applies to all channels
{
  "gateway": {
    "channelHealthRestartMode": "reconnect"  // "stop-start" | "reconnect" (default: "stop-start")
  }
}

// Per-channel override
{
  "channels": {
    "discord": { "healthMonitor": { "restartMode": "reconnect" } },
    "telegram": { "healthMonitor": { "restartMode": "reconnect" } }
  }
}

• "stop-start" (default): existing behavior — full teardown + restart, resets attempt counter
• "reconnect": lightweight recovery — stop + start without resetting the restart counter, signalling to the channel runtime that this is a soft reconnect. Falls back to "stop-start" on error.

Changes

Backwards Compatibility

Default is "stop-start" — no behaviour change for existing installs. Opt-in to "reconnect" explicitly.