Fix Nemotron Ultra Telegram replies by disabling compact tool search.

Set toolSearch false for Ultra and document web_search usage so Tavily runs instead of failing tool_search_code.

Author: maggiezha

nemoclaw-blueprint/model-specific-setup/openclaw/nemotron-ultra-managed-inference.json

@@ -0,0 +1,17 @@
+{
+  "$schema": "../schema.json",
+  "id": "nemotron-ultra-managed-inference",
+  "agent": "openclaw",
+  "description": "Disable OpenClaw compact tool-search for Nemotron Ultra; the model invents invalid tool_search_code instead of calling web_search directly.",
+  "match": {
+    "modelIds": ["nvidia/nvidia/llama-3.1-nemotron-ultra-253b-v1"],
+    "providerKey": "inference",
+    "inferenceApi": "openai-completions",
+    "baseUrl": "https://inference.local/v1"
+  },
+  "effects": {
+    "openclawTools": {
+      "toolSearch": false
+    }
+  }
+}

scripts/generate-openclaw-config.py

@@ -885,6 +885,8 @@ def _append_web_search_answer_instruction(config: dict) -> None:
         "\n\n## Web search (NemoClaw)\n"
         f"When you use web search to answer a question, tell the user clearly: "
         f"**{usage_line}** before summarizing search results.\n"
+        "Call the **`web_search`** tool with a `query` argument. "
+        "Do **not** use `tool_search_code` or invent `openclaw.tools.*` APIs.\n"
     )
     agents_md.parent.mkdir(parents=True, exist_ok=True)
     existing = ""