[BUG]: outputSchema validation fails for upstream servers returning structuredContent (v1.0.0-RC2)

[btbinaryjanitor]

Description

Tool calls through ContextForge virtual servers (Streamable HTTP transport) fail with "Output validation error: outputSchema defined but no structured output returned" when the upstream MCP server returns both content and structuredContent. This appears to be a regression or incomplete fix of #1406.

Affected Version

v1.0.0-RC2 (deployed on Cloud Run)

Steps to Reproduce

1. Register a Google BigQuery MCP gateway:

   • URL: https://bigquery.googleapis.com/mcp
   • Transport: STREAMABLEHTTP
   • Auth: OAuth client_credentials (via GCP token proxy)

2. BigQuery MCP tools define outputSchema (large JSON Schema for table/dataset metadata)

3. Call any tool through a ContextForge virtual server:

   {
     "jsonrpc": "2.0", "id": 2,
     "method": "tools/call",
     "params": {
       "name": "google-bigquery-list-dataset-ids",
       "arguments": {"projectId": "my-project"}
     }
   }

4. Response:

   {
     "jsonrpc": "2.0", "id": 2,
     "result": {
       "content": [{"type": "text", "text": "Output validation error: outputSchema defined but no structured output returned"}],
       "isError": true
     }
   }

Expected Behavior

The tool call should succeed, passing through the upstream response including structuredContent.

Direct Call Works

Calling the same tool directly on https://bigquery.googleapis.com/mcp (bypassing ContextForge) returns correctly:

{
  "result": {
    "content": [{"text": "{}", "type": "text"}],
    "structuredContent": {}
  }
}

The upstream server IS returning structuredContent, but ContextForge strips it before validation, then fails the validation check.

Also Affects

• Google Cloud Run MCP (https://run.googleapis.com/mcp) — all 5 tools define outputSchema
• Likely any upstream MCP server that defines outputSchema on its tools

Relation to #1406

#1406 fixed streamablehttp_transport.py's call_tool to return structured content (merged via PR #1412, released in BETA-1). The error message is slightly different:

• Ours: "Output validation error: outputSchema defined but no structured output returned"
• [BUG]: Missing structured content for virtual server in streamable HTTP #1406: "Tool has an output schema but did not return structured content"

This suggests either a regression in RC2, or a second validation checkpoint that wasn't covered by the original fix.

Environment

• ContextForge v1.0.0-RC2 on Google Cloud Run
• Upstream: Google BigQuery MCP (bigquery.googleapis.com/mcp) and Google Cloud Run MCP (run.googleapis.com/mcp)
• Transport: Streamable HTTP
• Virtual server with tools from multiple gateways

[bogdanmariusc10]

Thanks for reporting this, @btbinaryjanitor !

I've investigated this issue and have some findings to share:

The error message doesn't exist in the current codebase.

The specific error "Output validation error: outputSchema defined but no structured output returned" isn't present in v1.0.0-RC-2. I searched the entire codebase and found no matches.

How the code currently handles this scenario:

The validation logic in tool_service.py:1136-1137 explicitly returns True (success) when no structured content is found, even if outputSchema is defined. The code only validates structured content when it's actually present - it doesn't fail when outputSchema exists but structuredContent is missing.

Additionally:

• The streamable HTTP transport properly extracts structuredContent from upstream responses
• The tool service accepts both structuredContent and structured_content field names

Possible explanations:

1. Upstream server behavior - The error might originate from Google's BigQuery/Cloud Run MCP servers rather than ContextForge
2. Client-side validation - Your MCP client might be performing this validation
3. Configuration - Worth checking the virtual server and gateway configuration
4. Deployment mismatch - The current main version includes significant improvements and fixes over the RC-2 version

Next steps to diagnose:

Would you be able to enable debug logging (LOG_LEVEL=DEBUG) and share the full request/response flow? That would help us see exactly where this error is coming from. Also worth testing with our included output_schema_test_server to isolate whether it's specific to the Google MCP servers.

Let me know what you find!

[btbinaryjanitor]

Update: Confirmed ContextForge is relaying correctly — this is a client-side issue

Thanks @bogdanmariusc10 for the investigation. After thorough testing, you're right — ContextForge is not the problem here. The gateway is relaying both content and structuredContent correctly. The issue is on the client side (Claude Code's MCP client rendering).

Test Environment

• ContextForge image: ghcr.io/ibm/mcp-context-forge:latest (sha256:70a3a22942779c44..., pushed 2026-03-31)
• Upstream MCP server: run.googleapis.com/mcp (Google Cloud Run MCP)
• Auth: OAuth via token proxy (SA-based, client_credentials)
• Transport: Streamable HTTP

Evidence

1. Direct call to Google MCP (bypassing ContextForge):

POST https://run.googleapis.com/mcp
{"method": "tools/call", "params": {"name": "list_services", ...}}

Result:
  Has content:           True  (1 text item, 2021 chars)
  Has structuredContent: True  (5 services)

2. Tool definition confirms outputSchema is defined:

Tool: list_services
Has outputSchema: True
outputSchema type: object
outputSchema properties: [project, services]

All 5 Cloud Run MCP tools have outputSchema defined. Google returns responses with both content (text) and structuredContent (JSON) simultaneously.

3. Call through ContextForge gateway MCP endpoint:

POST https://mcpgateway.buildertrend.com/gateways/{gateway_id}/mcp
{"method": "tools/call", "params": {"name": "google-cloud-run-list-services", ...}}

Result (4274 bytes):
  Has content: True (1 text item, 2021 chars)
  Has structuredContent: True (5 services)
  isError: False

4. Call through ContextForge virtual server MCP endpoint:

POST https://mcpgateway.buildertrend.com/servers/{server_id}/mcp
{"method": "tools/call", "params": {"name": "google-cloud-run-list-services", ...}}

Result (4274 bytes):
  Has content: True (1 text item, 2021 chars)
  Has structuredContent: True (5 services)
  isError: False

5. Claude Code MCP client displays:

(mcp__contextforge-all-tools__google-cloud-run-list-services completed with no output)

Conclusion

ContextForge relays the full response including both content and structuredContent faithfully. The "no output" is rendered by Claude Code's MCP client, which appears to not handle responses containing structuredContent (or possibly both content + structuredContent simultaneously).

This is a client-side rendering issue, not a ContextForge bug. I'll investigate on the Claude Code / MCP SDK side.

Closing the ContextForge side of this — appreciate the quick response and the pointers to the validation code. The gateway is working as expected.