MCP server: dashboard tool schemas use $ref as a property name, breaking Kimi/Moonshot API clients (400 invalid schema)

Describe the bug

The hosted Logfire MCP server (https://logfire-us.pydantic.dev/mcp) exposes tool schemas that are rejected by the Moonshot/Kimi chat completions API (e.g. model kimi-for-coding via the kimi-code provider) with HTTP 400, before any tool call happens:

400 tools.function.parameters is not a valid moonshot flavored json schema, details: <At path 'properties.definition.anyOf.properties.spec.properties.layouts.items.properties.spec.properties.items.items.properties.content.properties.$ref': $ref must be a string>

The whole /chat/completions request is rejected because the tools payload contains one invalid schema. This makes the Logfire MCP server unusable from any client backed by the Kimi/Moonshot API (e.g. coding agents dispatching sub-agents on kimi-for-coding), while clients using other providers (Anthropic, OpenAI, DeepSeek) work fine.

Root cause

The dashboard_create and dashboard_update tools accept a Perses-compatible dashboard definition. Inside that schema, the grid layout item content object uses $ref as a plain property name (Perses panels are referenced as {"$ref": "#/spec/panels/..."}):

"content": {
  "properties": {
    "$ref": { "type": "string" }
  },
  "required": ["$ref"],
  "type": "object"
}

This is valid JSON Schema (draft 2020-12) — $ref here is just a data field name. However, the Moonshot/Kimi API validator treats the property name $ref as the JSON Schema keyword $ref (whose value must be a string JSON Pointer), sees an object instead, and rejects the entire tools payload.

Expected behavior

The MCP server should not emit schemas that break stricter JSON Schema validators. Possible fixes:

1. Avoid $ref as a property name in the tool input schema (e.g. rename the field, or describe panel references with a different keyword).
2. Or adjust the emitted schema so strict validators that special-case $ref do not reject it.

Environment

• MCP endpoint: https://logfire-us.pydantic.dev/mcp (remote hosted server)
• Client: oh-my-pi coding harness dispatching sub-agents on kimi-code / kimi-for-coding
• Affected tools: dashboard_create, dashboard_update (also surfaced as mcp__logfire_dashboard_create / mcp__logfire_dashboard_update)
• Related: https://github.com/pydantic/logfire-mcp/issues/37 (archived repo) reported a similar "JSON schema is invalid" rejection from Claude Code.

Start at the hosted MCP endpoint and inspect how the dashboard_create and dashboard_update tools generate schemas for the Perses-compatible definition input. Reproduce the tools payload against a Moonshot/Kimi validator, then verify that the emitted schema is accepted without breaking clients that currently work. Done means both affected tool schemas no longer trigger the reported invalid-$ref error.