MCP tool schemas are not sanitized for Anthropic: root-level anyOf/oneOf/allOf 400s, and tool.definition never sees MCP tools

Description

Any MCP server exposing a tool whose inputSchema has a root-level anyOf / oneOf / allOf makes every request to an Anthropic-backed model fail with a 400, before a single tool is called:

AI_APICallError: tools.5.custom.input_schema: input_schema does not support oneOf, allOf, or anyOf at the top level

ProviderTransform.schema already normalises tool schemas for OpenAI/Azure, Moonshot/Kimi and Gemini, but has no Anthropic branch, so the schema reaches @ai-sdk/anthropic / @ai-sdk/github-copilot verbatim. Anthropic accepts these combinators nested inside a property — only the root is rejected — so folding them into the root object schema is enough to fix it.

There is a second, separable half: MCP tools never reach the tool.definition hook, so a plugin cannot work around this either. With a plugin that only logs input.toolID, a run whose request carried 18 MCP tools plus the built-ins fired the hook 12 times, all built-in:

bash edit glob grep invalid question read skill task todowrite webfetch write

Built-in and plugin tools go through ToolRegistry.tools, which triggers tool.definition. MCP tools are converted separately in the session tool-assembly path and skip it. That leaves proxying the MCP server and rewriting tools/list as the only user-space fix, which is what I ended up doing.

Steps to reproduce

1. repro-mcp.mjs — dependency-free stdio MCP server with one offending tool:

let buffer = "";
const send = (message) => process.stdout.write(JSON.stringify(message) + "\n");

process.stdin.on("data", (chunk) => {
  buffer += chunk;
  let newline;
  while ((newline = buffer.indexOf("\n")) >= 0) {
    const line = buffer.slice(0, newline).trim();
    buffer = buffer.slice(newline + 1);
    if (!line) continue;
    const message = JSON.parse(line);
    if (message.method === "initialize") {
      send({ jsonrpc: "2.0", id: message.id, result: {
        protocolVersion: "2024-11-05",
        capabilities: { tools: {} },
        serverInfo: { name: "repro", version: "1.0.0" },
      }});
    } else if (message.method === "tools/list") {
      send({ jsonrpc: "2.0", id: message.id, result: { tools: [{
        name: "pick_one",
        description: "Pass exactly one of a or b.",
        inputSchema: {
          type: "object",
          properties: { a: { type: "string" }, b: { type: "string" } },
          anyOf: [{ required: ["a"] }, { required: ["b"] }],
        },
      }]}});
    } else if (message.method === "tools/call") {
      send({ jsonrpc: "2.0", id: message.id, result: { content: [{ type: "text", text: "ok" }] } });
    } else if (message.id !== undefined) {
      send({ jsonrpc: "2.0", id: message.id, result: {} });
    }
  }
});

2. opencode.json beside it:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "repro": { "type": "local", "command": ["node", "repro-mcp.mjs"], "enabled": true }
  }
}

3. opencode run --pure -m github-copilot/claude-haiku-4.5 "reply only: ok" → the 400 above.
   opencode run --pure -m github-copilot/gpt-5-mini "reply only: ok" → ok. Same config, same server; it is specific to the Anthropic-family providers.

Real-world case that led me here: the ai-memory MCP server ships one such tool, which takes down every Claude session in OpenCode while GPT and Gemini sessions are fine.

What I would expect

1. ProviderTransform.schema flattens root-level anyOf / oneOf / allOf into the root object schema for @ai-sdk/anthropic, @ai-sdk/google-vertex/anthropic and @ai-sdk/github-copilot, as it already does for the other providers.
2. MCP tool definitions also run through tool.definition, so plugins can patch provider quirks without proxying the server.

Plugins

None — reproduced with --pure.

OpenCode version

1.18.25

Operating System

Arch Linux (CachyOS), x86_64

Terminal

N/A — reproduced with non-interactive opencode run.