Ollama Cloud Does Not Support custom_tool_call Payload Used by Copilot CLI #3839

Summary

When using Copilot CLI v1.0.63 in Fleet Mode with a BYOK model routed through Ollama Cloud, requests may fail with the following error:

400 litellm.BadRequestError: OpenAIException - {"error":{"message":"input[69]: unknown input item type: \"custom_tool_call\"","type":"invalid_request_error","param":null,"code":null}}

Received Model Group=glm-5.2:cloud
Available Model Group Fallbacks=None

The failure does not appear to be model-specific.

Instead, the request generated by Copilot CLI Fleet Mode contains a custom_tool_call input item type that Ollama Cloud currently does not support.

Environment

• Copilot CLI
• Fleet Mode enabled
• BYOK Provider: Ollama Cloud
• Gateway: LiteLLM
• Example Model: glm-5.2:cloud

Actual Behavior

When Fleet Mode invokes tools or agent workflows, Copilot CLI sends one or more custom_tool_call items within the request payload.

Ollama Cloud rejects the request with:

unknown input item type: "custom_tool_call"

The request is rejected before reaching model inference.

Expected Behavior

One of the following should happen:

• Ollama Cloud supports the custom_tool_call input type used by Fleet Mode.
• Copilot CLI detects that the target provider does not support custom_tool_call and falls back to a compatible format.
• Copilot CLI provides a configuration option to disable Fleet-specific tool payloads for OpenAI-compatible providers with limited feature support.

Reproduction

1. Configure Copilot CLI to use a BYOK model through LiteLLM.
2. Configure LiteLLM to route requests to Ollama Cloud.
3. Enable Fleet Mode.
4. Execute a prompt that triggers agent or tool execution.
5. Observe the following error:

400 litellm.BadRequestError: OpenAIException - {"error":{"message":"input[69]: unknown input item type: \"custom_tool_call\"","type":"invalid_request_error","param":null,"code":null}}

Notes

Based on the error message, the request appears to be rejected by the Ollama Cloud API compatibility layer before reaching the underlying model.

This suggests the issue is related to Fleet Mode request formatting and provider capability compatibility rather than the specific model being used.

Any model hosted through Ollama Cloud may be affected when Fleet Mode emits custom_tool_call payloads.