[Batch API] Remote MCP (tools.mcp_servers) + structured output → every row fails code: 3 INVALID_ARGUMENT, yet job reports BATCH_STATE_SUCCEEDED

[Ahmed-IP-01]

[Batch API] Remote MCP (tools.mcp_servers) + structured output → every row fails code: 3 INVALID_ARGUMENT, yet job reports BATCH_STATE_SUCCEEDED

Description

In Batch mode, a request combining a remote MCP server (tools.mcp_servers) with a
structured-output directive (response_mime_type: "application/json" + response_json_schema,
or response_schema) is accepted, runs, and finishes as BATCH_STATE_SUCCEEDED — but every
request inside fails with code: 3, "Request contains an invalid argument."
(failedRequestCount == requestCount). No error at submission; you only find out after downloading
the responses. Dropping the structured-output directive makes the same MCP request succeed, so the
input shape is valid — it's the mcp_servers + structured-output combination.

Environment

• google-genai 2.8.0 · Python 3.11.12 · macOS 15
• Gemini Developer API, models/{model}:batchGenerateContent — reproduced on both v1beta and v1alpha
• Model gemini-3-flash-preview (also 2.5-flash/pro, 3.5-flash, 3.1-pro-preview)

Input (one batch request line)

{"request": {"contents": [{"role": "user", "parts": [{"text": "Find Apple's HQ city; answer JSON {hq}."}]}], "tools": [{"mcp_servers": [{"name": "web", "streamable_http_transport": {"url": "https://<mcp-host>/mcp/web", "headers": {"x-api-key": "<redacted>"}}}]}], "generationConfig": {"responseMimeType": "application/json", "responseJsonSchema": {"type": "object", "properties": {"hq": {"type": "string"}}, "required": ["hq"]}}}}

Submitted via POST /v1beta/models/gemini-3-flash-preview:batchGenerateContent with body
{"batch": {"inputConfig": {"requests": {"requests": [ <the object above>, ... ]}}}}.

Actual output (real batch result)

{
  "metadata": {
    "state": "BATCH_STATE_SUCCEEDED",
    "batchStats": { "requestCount": "2", "failedRequestCount": "2" },
    "output": { "inlinedResponses": { "inlinedResponses": [
      { "error": { "code": 3, "message": "Request contains an invalid argument." } },
      { "error": { "code": 3, "message": "Request contains an invalid argument." } }
    ]}}
  },
  "done": true
}

(On gemini-2.5 the message is explicit: "Function calling with a response mime type: 'application/json' is unsupported". Changing the MIME type — xml/yaml/enum — doesn't help; only
text/plain/no-schema is accepted.)

Expected

Either return schema-valid JSON per row, or reject at submission with a clear message, and
do not report BATCH_STATE_SUCCEEDED when failedRequestCount == requestCount.

[Venkaiahbabuneelam]

Hi @Ahmed-IP-01, Thanks for reaching out us!
To assist us in effectively investigating and reproducing it, please provide a minimal, runnable code example, including the complete error logs.

Thanks

[Ahmed-IP-01]

Remote MCP + response_json_schema: opaque INVALID_ARGUMENT (Batch: code 3) when an MCP tool's inputSchema contains a property with both anyOf and a sibling type

Thanks! Minimal repro + logs below.

Summary

Combining a remote MCP server (tools.mcp_servers) with structured output
(generationConfig.response_mime_type: "application/json" + response_json_schema) fails when
any MCP tool's inputSchema has a property that carries both anyOf and a redundant sibling
type (what Pydantic/FastMCP emit for Optional[...] fields), e.g.:

"tenant": { "anyOf": [{"type": "string"}, {"type": "null"}], "type": "string", "title": "tenant" }

All observations below are in Batch mode (models/{model}:batchGenerateContent):

• Without the response schema → batch succeeds, each row returns normally (MCP tools accepted as-is).
• With the response schema → the batch job still finishes BATCH_STATE_SUCCEEDED, but every
  row fails with {"error": {"code": 3, "message": "Request contains an invalid argument."}} and
  batchStats.failedRequestCount == requestCount. There is no error at submission and the message
  names no field/tool, so there is no signal until you download the outputs.
• Same batch against MCP servers with clean schemas (https://mcp.deepwiki.com/mcp,
  https://mcp.context7.com/mcp) → rows succeed and return schema-valid JSON. So remote MCP +
  structured output works in general; it's the tool-schema shape above that breaks it under
  structured output.

Environment

• google-genai 2.8.0 · Python 3.11.12 · macOS
• Gemini Developer API v1beta and v1alpha; models gemini-3-flash-preview, gemini-3.5-flash, gemini-3.1-pro-preview

Minimal runnable repro (Batch API)

Needs a reachable MCP server exposing a tool whose inputSchema contains the anyOf+sibling-type
property shown above (any FastMCP server with an Optional[...] parameter reproduces it). Then
submit a batch and poll it:

import json, time, urllib.request, urllib.error

KEY   = "<GEMINI_API_KEY>"
MODEL = "gemini-3-flash-preview"
ROOT  = "https://generativelanguage.googleapis.com/v1beta"
H     = {"x-goog-api-key": KEY, "Content-Type": "application/json"}

def req(method, path, body=None):
    r = urllib.request.Request(f"{ROOT}/{path}",
                               data=json.dumps(body).encode() if body else None,
                               headers=H, method=method)
    try:    return json.load(urllib.request.urlopen(r, timeout=60)), None
    except urllib.error.HTTPError as e: return None, json.loads(e.read().decode())

def one(text):
    return {"request": {
        "contents": [{"role": "user", "parts": [{"text": text}]}],
        "tools": [{"mcp_servers": [{
            "name": "srv",
            "streamable_http_transport": {
                "url": "https://<mcp-server-with-optional-param>/mcp",
                "headers": {"...": "..."}}}]}],
        # remove this generationConfig block and the SAME batch succeeds per row:
        "generationConfig": {"responseMimeType": "application/json",
                             "responseJsonSchema": {"type": "object",
                                 "properties": {"a": {"type": "string"}}, "required": ["a"]}},
    }}

body = {"batch": {"displayName": "mcp-schema-repro", "inputConfig": {"requests": {"requests": [
    one("Use the tool to find Apple's HQ city; answer JSON {a}."),
    one("Use the tool to find Netflix's HQ city; answer JSON {a}."),
]}}}}

created, err = req("POST", f"models/{MODEL}:batchGenerateContent", body)
assert not err, err
name = created["name"]
print("submitted:", name, "state:", created["metadata"]["state"])   # BATCH_STATE_PENDING

while True:
    time.sleep(15)
    got, _ = req("GET", name)
    print("state:", got["metadata"]["state"])
    if got.get("done"):
        print(json.dumps(got, indent=2))       # <-- output shown below
        break

Offending MCP tool inputSchema (as returned by the server's tools/list):

{
  "type": "object",
  "properties": {
    "queries": {"type": "array", "items": {"type": "string"}},
    "tenant":  {"anyOf": [{"type": "string"}, {"type": "null"}], "type": "string", "title": "tenant"},
    "mode":    {"anyOf": [{"type": "string", "enum": ["trusted","all"]}, {"type": "null"}], "type": "string", "default": "all"}
  },
  "required": ["queries"]
}

Complete error log (real batch run)

Console output from running the script above verbatim:

submitted: batches/e3drvvn52qtyec4yg3yw4gsh3k71ndog9nu4 state: BATCH_STATE_PENDING
state: BATCH_STATE_RUNNING
state: BATCH_STATE_RUNNING
state: BATCH_STATE_RUNNING
state: BATCH_STATE_RUNNING
state: BATCH_STATE_RUNNING
state: BATCH_STATE_RUNNING
state: BATCH_STATE_SUCCEEDED

Final polled job object (print(json.dumps(got, indent=2))) — job finishes
BATCH_STATE_SUCCEEDED, yet failedRequestCount == requestCount and every row is code: 3:

{
  "name": "batches/e3drvvn52qtyec4yg3yw4gsh3k71ndog9nu4",
  "metadata": {
    "@type": "type.googleapis.com/google.ai.generativelanguage.v1main.GenerateContentBatch",
    "model": "models/gemini-3-flash-preview",
    "displayName": "mcp-schema-repro",
    "output": {
      "inlinedResponses": {
        "inlinedResponses": [
          {
            "error": {
              "code": 3,
              "message": "Request contains an invalid argument."
            }
          },
          {
            "error": {
              "code": 3,
              "message": "Request contains an invalid argument."
            }
          }
        ]
      }
    },
    "createTime": "2026-07-10T05:05:40.067972628Z",
    "endTime": "2026-07-10T05:07:29.761773602Z",
    "updateTime": "2026-07-10T05:07:29.761773652Z",
    "batchStats": {
      "requestCount": "2",
      "failedRequestCount": "2"
    },
    "state": "BATCH_STATE_SUCCEEDED",
    "name": "batches/e3drvvn52qtyec4yg3yw4gsh3k71ndog9nu4"
  },
  "done": true,
  "response": {
    "@type": "type.googleapis.com/google.ai.generativelanguage.v1main.GenerateContentBatchOutput",
    "inlinedResponses": {
      "inlinedResponses": [
        {
          "error": {
            "code": 3,
            "message": "Request contains an invalid argument."
          }
        },
        {
          "error": {
            "code": 3,
            "message": "Request contains an invalid argument."
          }
        }
      ]
    }
  }
}

Asks

1. Accept common Pydantic/FastMCP MCP tool schemas (anyOf + sibling type) under structured
   output, as they already are without it — or
2. Return an actionable error naming the offending tool + field instead of a generic
   INVALID_ARGUMENT, and fail fast at batch submission rather than per-row at execution.
3. Don't report BATCH_STATE_SUCCEEDED when failedRequestCount == requestCount.