Run Flow component returns empty content when invoked as an Agent tool

[saranuh]

Bug Description

When a Run Flow component is invoked by an Agent as a tool, the tool receives the correct input argument but returns empty content.

The same Run Flow component works correctly when executed directly without tool mode.

A similar problem was reported in #14791. I added my reproduction details there, but because that issue is already closed and the problem remains reproducible on Langflow 1.12.1, I am opening this issue to document the current behaviour.

This may also be related to the tool-invocation problem addressed by #14793.

Reproduction

Minimal setup
Subflow

Chat Input → Chat Output

I also tested:

Text Input → Chat Output

The subflow only passes the received input directly to the output. It does not contain an Agent or language model.

Parent flow

Chat Input → Agent → Chat Output

The Run Flow component is connected to the Agent as a tool.

The Agent is instructed to call the Run Flow tool and pass its received input unchanged.

Steps to reproduce

1. Create a subflow containing only Chat Input → Chat Output.
2. Create a parent flow with an Agent.
3. Select the subflow in a Run Flow component.
4. Connect the Run Flow component to the Agent as a tool.
5. Send Test123 to the parent flow.
6. Inspect the Agent trace and Run Flow tool result.

Expected behavior

The Run Flow tool should pass Test123 to the subflow and return:

Test123

When I execute the Run Flow component directly and provide Test123 as its input, the subflow correctly returns Test123.

When the Agent invokes the same component as a tool, the tool-call arguments contain the correct value:

{
  "flow_tweak_data": {
    "TextInput-XXXX~input_value": "Test123"
  }
}

However, the tool returns empty content:

{
  "type": "tool_end",
  "output": {
    "content": ""
  }
}

Who can help?

@italojohnny @ogabrielluiz

Operating System

Linux Docker container on Windows 11

Langflow Version

1.12.1

Python Version

None

Screenshot

No response

Flow File

No response

[L4XB]

Traced this to two objects rather than one. PR #15105.

ComponentToolkit copies the component on every tool call, which is deliberate (#8791, so concurrent calls do not share state), applies the call's arguments to the copy, and then resolves the output method as getattr(comp, method_name, output_method) where output_method is bound to the component the toolkit was built from.

Run Flow's output methods are not on the class. _register_flow_output_method writes one per selected-flow output onto the instance, and Component.__deepcopy__ rebuilds the component through type(self)(**kwargs) rather than copying __dict__, so the copy has none of them. The getattr default then takes over, and the method that runs is bound to the original. Your flow_tweak_data went to the copy; the resolver reads self._attributes on the original, where it is still {}. The sub-flow runs with no tweak, its input node gets nothing, and the tool answers empty. Nothing raises, which is why the tool call looks correct in the trace and only the result is wrong.

Measured on release-1.12.2:

resolver survives the deepcopy: False
wrapper's local_method is bound to the ORIGINAL: True
tool arguments landed on the copy    : {'TextInput-1~input_value': 'hello'}
what the running method will read    : {}

That also explains the part of your report that looked contradictory: the tool-call arguments are correct because they really were delivered, just onto the object that then did not run.

The PR rebuilds the flow output methods on the copy. If you can run a build from it, the workflow from your report is the check I could not do here: Test123 should come back instead of empty content.

[andifilhohub]

LE-2650

[andifilhohub]

The fix has been merged.