Segmentation Fault When Recursion Limit Too High

[dOrgJelli]

Posting this for others who run into this issue. Langchain uses repr internally, which does not account for recursive object references (PR to fix this). This becomes a problem when derived Langchain classes (ex: CrewAgentExecutor) are serialized (ex: invoke -> dumpd) using repr.

This problem is "hidden" due to a low sys.getrecursionlimit() being set by default (1000). If you increase this number (100,000), a Segmentation Fault is encountered due to the process running out of memory. An example object-reference-recursion can be seen here: self.task.agent.agent_executor.task.agent.agent_executor.task.agent

To solve this problem, you can define a safe_repr implementation (see fix PR above) and simply override the global repr implementation (assuming langchain hasn't fixed this before then):

def patch_langchain():
    default_repr = __builtins__["repr"]

    def safe_repr(obj, seen=None):
        ...see-pr...
        else:
            # Handle general class instances
            if hasattr(obj, "__dict__"):
                result.append(f"<{obj.__class__.__name__}: ")
                for key, value in obj.__dict__.items():
                    result.append(f"{key}={safe_repr(value, seen=seen)}, ")
                result.append(">")
            else:
                # Fallback for objects without __dict__ or non-container types
                result.append(default_repr(obj))

        seen.remove(obj_id)
        return ''.join(result)

    __builtins__["repr"] = safe_repr

And you can simply call patch_langchain() inside your program's entry-point.

Hope this makes sense, please ask questions if it doesn't, happen to explain! Took us ~4 hours to get to the root of this segmentation fault, hope this saves people time.