[MCP] reinject JSON parse & runtime tool errors back into the chat history

[hanouticelina]

Copy-paste from the JS PR: huggingface/huggingface.js#1511 which I find useful.

this should prevents crashes when the model emits malformed JSON and gives the LLM immediate feedback it can reason over, matching the js implem.

[Copilot]

Pull Request Overview

This PR enhances process_single_turn_with_tools to catch and re-inject JSON parsing and tool execution errors back into the chat history, preventing crashes when malformed JSON is produced or tool calls fail.

• Add JSONDecodeError handler to wrap malformed JSON from the model and yield an error message
• Wrap tool execution in a try/except to catch runtime errors and report them as tool messages

Comments suppressed due to low confidence (2)

src/huggingface_hub/inference/_mcp/mcp_client.py:315

• Introduce unit tests covering the JSONDecodeError branch to verify that malformed JSON from the model is caught and properly yielded as a tool message.

except json.JSONDecodeError as err:

src/huggingface_hub/inference/_mcp/mcp_client.py:343

• [nitpick] Consider standardizing the error message format and avoid leaking raw exception text; for example, use a concise message and log the full exception separately.

tool_message["content"] = f"Error: MCP tool call failed with error message: {err}"

[Copilot]

Avoid catching the broad Exception; consider handling specific exceptions thrown by session.call_tool or re-raising unexpected errors to avoid masking serious issues.

Suggested change

	except Exception as err:
	tool_message["content"] = f"Error: MCP tool call failed with error message: {err}"
	except (ValueError, TypeError, json.JSONDecodeError) as err:
	tool_message["content"] = f"Error: MCP tool call failed with error message: {err}"
	except Exception as err:
	logger.error(f"Unexpected error during MCP tool call: {err}")
	raise

[Wauplin]

Agree in general but disagree on this specific use case. Let's keep it like this

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Wauplin]

Thanks!

[Wauplin]

Agree in general but disagree on this specific use case. Let's keep it like this