Summary

When a user drops an X.com (or any domain) URL, the agent reflexively calls web_fetch. For X.com this always fails — X blocks unauthenticated scrapers. The right path is to use the xread skill (Grok API), but no enforcement mechanism exists today. Text rules in AGENTS.md aren't reliable enough because pattern-matching fires before the rule is read.

The before_tool_call hook already exists and is the perfect enforcement point.

Proposed: URL routing config in openclaw.json

{
  "tools": {
    "web": {
      "fetch": {
        "urlRouting": [
          {
            "match": "x\.com|twitter\.com",
            "action": "redirect",
            "redirectTo": "skill:xread",
            "reason": "X.com blocks unauthenticated fetch. Use xread skill (Grok API) instead."
          },
          {
            "match": "linkedin\.com",
            "action": "warn",
            "reason": "LinkedIn blocks scrapers — results may be empty."
          }
        ]
      }
    }
  }
}

How it would work

The before_tool_call hook already receives:

PluginHookBeforeToolCallEvent = {
  toolName: string;   // e.g. "web_fetch"
  params: Record<string, unknown>;  // includes { url: "https://x.com/..." }
}

And can return:

PluginHookBeforeToolCallResult = {
  block?: boolean;
  blockReason?: string;
  params?: Record<string, unknown>;  // mutate params (e.g. rewrite URL)
  requireApproval?: { ... };
}

The routing logic would live in a bundled plugin that:

1. Reads tools.web.fetch.urlRouting from config
2. On before_tool_call where toolName === "web_fetch":
   • Extracts the url param
   • Checks against routing rules (regex match on domain)
   • If match + action is redirect: blocks the tool call and injects a system note telling the agent to use the specified skill/tool instead
   • If match + action is warn: allows but prepends a warning to the tool result

Why this matters beyond x.com

This generalizes to any domain-specific routing:

• Paywalled sites → redirect to a summarize skill
• Internal tools → redirect to an MCP server
• Rate-limited APIs → warn + suggest caching
• Security policy → block specific domains entirely

Agents today have no structured way to express "when you see a URL from domain X, use capability Y." This fills that gap with a config-driven, code-enforced approach rather than relying on text rules that can be skipped.

Implementation sketch

// src/plugins/url-routing.plugin.ts

import type { OpenClawPlugin } from './types.js';

export const urlRoutingPlugin: OpenClawPlugin = {
  id: 'core:url-routing',
  hooks: [
    {
      hookName: 'before_tool_call',
      handler: async (event, ctx) => {
        if (event.toolName !== 'web_fetch') return;
        
        const url = event.params?.url as string | undefined;
        if (!url) return;

        const rules = resolveUrlRoutingRules(ctx); // reads from config
        
        for (const rule of rules) {
          if (!new RegExp(rule.match).test(url)) continue;
          
          if (rule.action === 'redirect') {
            return {
              block: true,
              blockReason: `URL routing: ${url} matches "${rule.match}" — use ${rule.redirectTo} instead. Reason: ${rule.reason}`,
            };
          }
          
          if (rule.action === 'warn') {
            // inject warning into result via params mutation or post-hook
          }
        }
      },
      priority: 100, // run early
    },
  ],
};

Related

• before_tool_call hook already exists in src/plugins/hook-types.ts
• Existing test coverage: src/plugins/hooks.before-tool-call.test.ts
• PluginHookBeforeToolCallResult.block + blockReason already supported

This is additive — no breaking changes to existing plugin or hook behavior. The routing table defaults to empty (no rules = no change in behavior).

Surfaced by a real agent (Bartok) that keeps reflexively calling web_fetch on x.com despite text rules saying not to. Code enforcement > text rules.