Azure AI Foundry Agents - Project-scoped /openai/v1/responses endpoint returns HTTP 431 for ALL requests

Question

Environment:

• Region: Central India (centralindia)
• Resource: Azure AI Services (Cognitive Services) - brsr-quant-pot-resource
• Project: brsr_quant_pot
• SDK: azure-ai-projects==2.0.1, openai==2.30.0
• Model deployment: gpt-4o-mini (gpt-4.1-mini) - Status: Succeeded

Problem: All agents deployed in our Azure AI Foundry project return "An error occurred while processing your request" in the Playground. When tested via SDK or REST API, the project-scoped responses endpoint returns HTTP 431 (Request Header Fields Too Large) for every single request — regardless of header size, auth method, or agent complexity.

Key findings from testing:

1. Agent management APIs work fine on the project-scoped endpoint (project.agents.list(), create_version(), delete() all succeed).
2. The project-scoped responses endpoint fails for ALL requests:

POST https://<Resource_id>.
.services.ai.azure.com/api/projects/brsr_quant_pot/openai/v1/responses
→ HTTP 431 Request Header Fields Too Large

3. Even with a tiny 84-byte api-key header (no Bearer token), it still returns 431.
4. Even with only ONE minimal agent (no tools, 40-character instructions), it returns 431.
5. The resource-level endpoint works fine:

POST https://<Resource_id>.services.ai.azure.com/openai/v1/responses
→ HTTP 200 (but doesn't support agent_reference)

6. The Azure Function Apps backing the agents are running and return correct data when called directly with the function key.

Steps to reproduce:

# Get API key
API_KEY=$(az cognitiveservices account keys list --name brsr-quant-pot-resource --resource-group rs-esg-emissions --query key1 -o tsv)

# This FAILS with 431 (project-scoped):
curl -X POST "https://<Resource_id>.services.ai.azure.com/api/projects/brsr_quant_pot/openai/v1/responses" \
  -H "api-key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":"Hello","model":"gpt-4o-mini"}'
# Returns: Request Header Fields Too Large (HTTP 431)

# This WORKS (resource-level):
curl -X POST "https://<Resource_id>.services.ai.azure.com/openai/v1/responses" \
  -H "api-key: $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":"Hello","model":"gpt-4o-mini"}'
# Returns: HTTP 200

What this rules out:

• NOT a header size issue (84-byte key still triggers 431)
• NOT an agent/spec complexity issue (1 minimal toolless agent still triggers 431)
• NOT a model deployment issue (model is Succeeded, works on resource-level endpoint)
• NOT a client SDK issue (raw curl also fails)

Conclusion: The project routing layer on the /openai/v1/responses path is broken for this resource/project. Other paths on the same project-scoped endpoint (agent management) work correctly. This blocks all agent Playground testing and SDK-based agent queries.

Subscription ID: <REDACTED at Support side>

Is this a known service issue in the Central India region? Is there a workaround to use agent_reference on the resource-level endpoint? Or does the project need to be recreated?