langgraph/how-tos/return-when-recursion-limit-hits/

[giscus[bot]]

langgraph/how-tos/return-when-recursion-limit-hits/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/return-when-recursion-limit-hits/?h=recursion

[abprime]

The step count keeps increasing every invoke. How can we reset or make the lastStepManager start from 1 again for every invoke. The recursion_limit is per invoke not for across the invoke calls.

Any solution for this? Or is this a issue that needs to be fixed?

[abprime]

This is from the documentation:
https://langchain-ai.github.io/langgraph/concepts/low_level/?h=recursion#recursion-limit
The recursion limit sets the maximum number of super-steps the graph can execute during a single execution

[antonioalegria]

I'm hitting the exact same issue - have you opened an issue with this?

The step count keeps increasing over multiple invocations, so this solution effectively only works for the first invocation. I haven't found a way to track the "steps since last invocation".

[abprime]

No. I haven’t raised an issue yet. Wanted to see if there is a solution for this problem.

[antonioalegria]

Any update on the issue above?

[abprime]

Raised a PR #2027.
not sure when it is getting merged.

[abprime]

Fixed in #2105

[Fickincool]

Thanks for this! I was reading through the PR and saw that this should work on the create_react_agent, but I am using version 0.2.61 and I am getting Recursion errors with no returning state. Does anyone has an example of how this works in this case? or how to reset the recursion counter after a response is generated?

[ace-builds]

Is there a Javascript example of this how-to?

[mrctito]

Hi!

In my case, the tool is executing correctly and returning the response, but the graph does not stop. Any tips?

[seungwonme]

hello.

I'm not sure if I'm answering this correctly, but RemainingSteps only subtracts 1 from the value you have each time you enter a node, it doesn't have the ability to stop the graph.
The benefit of using RemainingSteps is that you don't have to include the logic to decrement -1 at each node.
This may be a bad analogy, but I think it's similar to the benefit of introducing middleware on the server.

def router(state: State):
    # Force the agent to end
    if state["remaining_steps"] <= 2:
        return END
    if state["value"] == "end":
        return END
    else:
        return "action"