Description

The create_alert_rule and update_alert_rule tools emit bare boolean JSON Schema values (e.g., "model": true) in their tool definitions. While valid per the JSON Schema specification (true means "accept any value"), LLM providers like Fireworks AI reject these with a 500 internal_server_error.

Reproduction

1. Connect mcp-grafana to any LLM gateway that routes to Fireworks AI
2. Include create_alert_rule or update_alert_rule in the tool list
3. Send a chat completion request with 26+ tools (including the alert rule tools)
4. Fireworks returns: {"error":{"object":"error","type":"internal_server_error","code":"invalid_request_error","message":"server had an error while processing your request, please retry again after a brief wait"}}

With 25 tools (excluding alert rule tools): succeeds.
With just create_alert_rule alone: fails.

Root Cause

The alert rule params include Go interface{} fields (likely QueryData or the alert rule model). When reflected to JSON Schema via the jsonschema library, interface{} produces a bare true. Fireworks AI's schema parser does not handle bare boolean schemas.

Affected Tools

• create_alert_rule — has "model": true in schema
• update_alert_rule — same

Expected Behavior

Bare boolean schemas should be replaced with equivalent object schemas:

• true → {"type": "object"} (or more specific if the type is known)
• false → {"not": {}}

Workaround

We applied a sanitizeSchema() function in our MCP adapter layer that recursively converts bare booleans before passing schemas to the LLM provider. See: androidStern-personal/openclaw-mcp-adapter#9

Environment

• mcp-grafana: Docker image grafana/mcp-grafana (latest as of 2026-02-23)
• Grafana: 11.x with Editor-role service account
• LLM Provider: Fireworks AI (Kimi K2.5 model)
• Gateway: LiteLLM v1.81.13 + OpenClaw v2026.2.22