Grok 4.5: 351 tools fail with HTTP 400 instead of reporting the 350-tool limit

Type: Bug

Copilot CLI Version: 1.0.84-5
OS Version: Windows 11 x64
Model: grok-4.5

Summary

When the total number of advertised tools exceeds 350, requests to grok-4.5 fail with HTTP 400 Bad Request.

Copilot CLI does not enforce or report the tool count limit before making the API call.

Steps to Reproduce

1. Use a test MCP configuration with no other user/plugin MCP servers enabled, so they do not add to the count.
2. Save schema-repro.mjs, grok-350.json, and grok-351.json from the Reproducer files section below in the same directory. Node.js must be available.
3. From that directory, run these native Copilot commands:

copilot --model grok-4.5 --no-custom-instructions --disable-builtin-mcps --disable-mcp-server computer-use --additional-mcp-config "@grok-350.json" --available-tools "mcp:*" --allow-all-tools -p "Reply with exactly OK. Do not call tools."
copilot --model grok-4.5 --no-custom-instructions --disable-builtin-mcps --disable-mcp-server computer-use --additional-mcp-config "@grok-351.json" --available-tools "mcp:*" --allow-all-tools -p "Reply with exactly OK. Do not call tools."

The fixture tools have unique names and identical simple schemas. mcp:* excludes native CLI tools; the MCP-only precondition isolates the fixture's 350/351 count. There is no SDK dependency or shell helper.

Actual Behaviour

With 350 tools:

• Request succeeds with exit code 0.

With 351 tools:

• Request immediately fails with:

Execution failed: 400 Bad Request (Request ID: ...)

Comparison:

The debug tool-surface records confirm exactly 350 and 351 advertised tools for these two runs.

Expected Behaviour

Copilot CLI should track provider-specific limits:

1. Enforce a tool count budget before sending the request.
2. If tool count exceeds 350 for Grok, report an actionable error explaining that the active tool count (351) exceeds Grok's maximum limit of 350 tools.
3. Alternatively, defer tools via tool search rather than sending the entire catalog upfront.

Notes

• Tested with inert synthetic tools with identical simple schemas ({"type":"object","properties":{"value":{"type":"string"}}}).
• The issue is purely triggered by the total number of advertised tools, regardless of whether tools are actually invoked by the model.

Impact

Users with multiple MCP servers easily exceed 350 tools in aggregate (native + MCP tools), causing Grok to become completely unusable across the entire session with an unhelpful 400 error.

import { createInterface } from 'node:readline';

const mode = process.argv[2];
const count = Number(process.argv[3] ?? 1);
const allowedModes = ['array-enum-bad', 'array-enum-fixed', 'count'];

if (!allowedModes.includes(mode) ||
    !Number.isInteger(count) ||
    count < 1 ||
    count > 400) {
  throw new Error('Usage: node schema-repro.mjs array-enum-bad|array-enum-fixed|count [1..400]');
}

const levels = { type: ['array', 'null'], items: { type: 'integer' } };
if (mode === 'array-enum-bad') {
  levels.enum = [0, 1, 2, 3, 4, 25];
} else {
  levels.items.enum = [0, 1, 2, 3, 4, 25];
}

const tools = Array.from({ length: mode === 'count' ? count : 1 }, (_, index) => ({
  name: `probe_${String(index).padStart(3, '0')}`,
  description: 'Schema diagnostic only; performs no operation.',
  inputSchema: {
    type: 'object',
    properties: mode === 'count' ? { value: { type: 'string' } } : { levels },
  },
}));

for await (const line of createInterface({ input: process.stdin, crlfDelay: Infinity })) {
  if (!line.trim()) {
    continue;
  }
  const request = JSON.parse(line);
  if (request.id === undefined) {
    continue;
  }

  let result;
  if (request.method === 'initialize') {
    result = {
      protocolVersion: request.params.protocolVersion,
      capabilities: { tools: {} },
      serverInfo: { name: 'schema-repro', version: '1.0.0' },
    };
  } else if (request.method === 'tools/list') {
    result = { tools };
  } else if (request.method === 'tools/call') {
    result = {
      isError: true,
      content: [{ type: 'text', text: 'This diagnostic never executes an operation.' }],
    };
  } else if (request.method === 'ping') {
    result = {};
  } else {
    process.stdout.write(`${JSON.stringify({
      jsonrpc: '2.0',
      id: request.id,
      error: { code: -32601, message: 'Method not found' },
    })}\n`);
    continue;
  }

  process.stdout.write(`${JSON.stringify({ jsonrpc: '2.0', id: request.id, result })}\n`);
}

{
  "mcpServers": {
    "schema_repro": {
      "type": "stdio",
      "command": "node",
      "args": ["schema-repro.mjs", "count", "350"],
      "tools": ["*"]
    }
  }
}

{
  "mcpServers": {
    "schema_repro": {
      "type": "stdio",
      "command": "node",
      "args": ["schema-repro.mjs", "count", "351"],
      "tools": ["*"]
    }
  }
}

Start by running the two native Copilot commands with grok-350.json and grok-351.json, using schema-repro.mjs to confirm the isolated tool counts. Trace the CLI request path that builds the advertised tool list and handles the Grok response; done means 350 tools still succeed and 351 produces an actionable limit error before the HTTP request.