litellm - 💡(How to fix) Fix Responses API → Chat Completions transformation silently passes through unsupported built-in tool types [1 pull requests]

Problem

When LiteLLM converts an OpenAI Responses API request to the Chat Completions API, several built-in tool types that have no valid equivalent in the Chat Completions spec are silently passed through to the provider. This causes provider-side 400 errors such as:

"tools[N].type:type is illegal"

Affected tool types

• local_shell — Codex CLI built-in shell tool (no Chat Completions equivalent)
• file_search — OpenAI file search tool (handled separately by emulated handler, but still leaks through in transformation)
• image_generation — OpenAI image generation tool
• namespace — Codex CLI 0.130.0 tool namespace
• mcp — MCP server tools are Responses-API-specific
• code_interpreter — Has no native Chat Completions equivalent, but previously passed through raw
• computer_use / computer_use_preview — Passed through raw, causing schema validation failures on generic providers
• Unknown / passthrough types (e.g. tool_search, code_execution_20250825, Vertex AI native tools) — silently forwarded

Repro

from litellm.responses.litellm_completion_transformation.transformation import (
    LiteLLMCompletionResponsesConfig,
)

# This should be dropped, but was previously passed through
tools = [{"type": "local_shell"}]
result, _ = LiteLLMCompletionResponsesConfig.transform_responses_api_tools_to_chat_completion_tools(
    tools=tools
)
print(result)  # Before fix: [{"type": "local_shell"}]  → provider errors

Proposed Fix

PR #27652 addresses this by:

1. Declaring unsupported built-in types via a class-level unsupported_tool_types constant:

   unsupported_tool_types = [
       "local_shell",
       "file_search",
       "image_generation",
       "namespace",
   ]

2. Converting code_interpreter to a function tool so the model can still generate code via a named tool call.

3. Mapping computer_use per provider:

   • Anthropic → preserves computer_use_preview type with function.parameters wrapper
   • Gemini/Vertex AI → flat dict with type="computer_use"
   • Generic providers → converted to a function tool

4. Dropping MCP tools with a warning log instead of passing them through.

5. Dropping unknown tool types with a warning log instead of passing them through.

6. Adding comprehensive test coverage for all of the above.

Related

• PR #27652 — fix implementation

Environment

• LiteLLM version: main (post 1.79.1)
• Python: 3.12+