Fix Ollama 5-Minute Timeout

Your Ollama model generates long responses fine via curl, but OpenClaw cuts the connection at exactly 5 minutes with a 500 error — then enters an infinite retry loop. The root cause is a hardcoded timeout in Node.js's HTTP client.

Symptoms

Confirm with Log Filtering

# Look for the specific undici error codes
openclaw logs --follow | grep -E 'UND_ERR_HEADERS_TIMEOUT|UND_ERR_BODY_TIMEOUT'

If you see UND_ERR_HEADERS_TIMEOUT, this is your problem.

Root Cause: undici 300-Second Default

OpenClaw uses Node.js's built-in undici HTTP client. This client has a hardcoded 300-second (5-minute) headersTimeout. If the server doesn't complete the response within this window, undici kills the connection.

The Fix: Monkey-Patch Preload Script

The solution is a Node.js preload script that sets the timeout and prevents other libraries from resetting it.

Step 1: Create the Preload Script

Save this as undici-timeout-preload.cjs (e.g., in your home directory):

const undici = require("undici");
const { EnvHttpProxyAgent } = undici;

const OPTS = {
  headersTimeout: 30 * 60 * 1000, // 30 minutes
  bodyTimeout: 0,                 // Disable body timeout
};

const realSet = undici.setGlobalDispatcher.bind(undici);

function enforce() {
  realSet(new EnvHttpProxyAgent(OPTS));
}

// 1. Set the custom dispatcher immediately
enforce();

// 2. Prevent other libraries from resetting it
undici.setGlobalDispatcher = function () {
  enforce();
};

Step 2: Load the Script at Startup

systemctl edit --user openclaw

# Add these lines:
[Service]
Environment="NODE_OPTIONS=--require /home/user/undici-timeout-preload.cjs"

NODE_OPTIONS="--require /path/to/undici-timeout-preload.cjs" \
  openclaw gateway start

$env:NODE_OPTIONS="--require C:\Users\YourName\undici-timeout-preload.cjs"
openclaw gateway start

Disable the Retry Loop

Even after fixing the timeout, you should prevent OpenClaw from auto-retrying expensive long generations:

# Set retry attempts to 1 (no retry) for Ollama
openclaw config set models.providers.ollama.retry '{"attempts": 1}' --json
openclaw gateway restart

This prevents OpenClaw from restarting the entire generation from scratch if anything goes wrong. You'll get an error message instead of silently burning GPU time in a loop.

Why Streaming Doesn't Help

You might think streaming would bypass this (since the first byte arrives quickly). However:

Verification

After applying the fix, verify it works:

# 1. Start gateway with the preload script
NODE_OPTIONS="--require /path/to/undici-timeout-preload.cjs" \
  openclaw gateway run --verbose

# 2. Send a prompt that takes >5 minutes
# 3. Monitor for timeout errors
openclaw logs --follow | grep -E 'UND_ERR|timeout|500'

If no UND_ERR_HEADERS_TIMEOUT appears and the response completes, the fix is working.

Still Stuck?

Run our diagnostic tool to catch issues automatically: