openclaw - 💡(How to fix) Fix Feature Request: Configurable streaming watchdog timeout threshold [4 comments, 5 participants]

Problem Description

When using models that perform extended reasoning/thinking (e.g., kimi-k2.5, DeepSeek-R1), the streaming watchdog frequently triggers warnings:

streaming watchdog: no stream updates for 30s; resetting status. 
The backend may have dropped this run silently — send a new message to resync.

This 30-second threshold appears to be hardcoded in the client UI and cannot be configured server-side.

Impact

1. False positives during legitimate reasoning: Models like kimi-k2.5 may legitimately spend 30-60+ seconds in deep thinking phases before producing output. The watchdog warning creates unnecessary alarm for users.

2. User experience disruption: The warning message implies something went wrong, prompting users to "resync" by sending new messages, which can:

   • Interrupt ongoing reasoning processes
   • Cause duplicate queries
   • Break conversation context

3. No server-side mitigation available: Despite configuring:

   • agents.defaults.timeoutSeconds: 1200 (20 min agent timeout)
   • reasoningDefault: "stream" (stream reasoning output)
   • blockStreamingChunk.minChars: 20 (more frequent chunking)

   The 30-second client watchdog still triggers independently.

Proposed Solution

Add a configurable streaming watchdog timeout in openclaw.json:

{
  "gateway": {
    "streamingWatchdog": {
      "noOutputTimeoutMs": 60000,  // configurable threshold, default 30000
      "enabled": true               // allow disabling for trusted setups
    }
  }
}

Alternatively, expose this via ACP stream configuration:

{
  "acp": {
    "stream": {
      "watchdogNoOutputMs": 60000
    }
  }
}

Use Cases

1. Extended reasoning models: Users deploying models with long thinking phases (30-120 seconds) can adjust thresholds appropriately.

2. Local/trusted setups: For personal assistant deployments on loopback-only gateways, users may prefer longer thresholds or disabling the watchdog entirely.

3. Network-aware configuration: Users with known network latency characteristics can tune thresholds based on their environment.

Current Workarounds (Limited Effectiveness)

• reasoningDefault: "stream" — Helps but requires models to emit reasoning tokens
• blockStreamingChunk.minChars — More frequent chunks but doesn't prevent legitimate thinking silence
• No actual watchdog timeout configuration exists

Related

The acp.stream.coalesceIdleMs parameter exists but controls text block coalescing, not the watchdog timeout mechanism.

Thank you for considering this feature request. A configurable watchdog threshold would significantly improve the experience for users working with reasoning-capable models.