Problem

When a channel (Discord, Telegram, etc.) goes stale, the OpenClaw health-monitor correctly restarts only the channel connector (via stopChannel + startChannel). However, this still invalidates the Anthropic prompt cache, which is process-level.

Real-world impact

• Normal operation: ~500 new tokens per turn (~99% cache hit rate)
• After any channel restart: ~120,000 tokens sent in full (0% cache hit for 2-3 turns)
• Cost difference per restart: ~240x token usage spike

Incident log (2026-03-30)

12:35:45 — health-monitor: restarting (reason: stale-socket)  [discord:default]
12:35:46 — discord gateway: WebSocket closed with code 1005
13:22:45 — discord gateway: Attempting resume with backoff 1000ms

The channel reconnect itself is fine — Discord handles it gracefully. The problem is that the cache warm-up after reconnect costs a full context re-send.

Root Cause

Anthropic's prompt cache is tied to the model inference session context on the Anthropic API side. Any interruption in the streaming connection — including a channel-level reconnect — forces a cold cache start on the next message. The agent's system prompt (~120k tokens) has to be re-sent in full.

Proposed Solution

Option A — Reconnect without stopping (preferred)

Instead of stopChannel → startChannel, try a lighter reconnect() path when the reason is stale-socket. Discord's own library already supports client.ws.reconnect() — this resumes the WebSocket session without destroying the client, which means the next Anthropic call can potentially reuse the cache warmup from before the disconnect.

Option B — Cache pre-warming after restart

After a channel restart, immediately send a lightweight keep-alive message to the active agent sessions to re-warm the Anthropic cache before the user sends their next message.

Option C — Config flag

Add gateway.channelStaleRestartMode:

{
  "gateway": {
    "channelStaleRestartMode": "reconnect" // "reconnect" | "stop-start" (default: "stop-start" for compat)
  }
}

Code Location

The relevant logic is in:

• src/gateway/channel-health-monitor.ts — runCheck() calls stopChannel + startChannel
• src/gateway/server-channels.ts — ChannelManager interface

Current Workaround

Disabling the health-monitor per-channel avoids the cache bust, since channels reconnect themselves:

{
  "channels": {
    "discord": { "healthMonitor": { "enabled": false } }
  }
}

This works because Discord (and most providers) implement their own reconnect logic. The health-monitor restart is only needed when a channel is truly dead (not just momentarily disconnected).

Why This Matters

With 25k+ OpenClaw users running Anthropic models with long system prompts, every stale-socket event silently multiplies token costs by ~240x for 2-3 turns. This is invisible to users and looks like random cost spikes.