fix(langchain): inline aafter_model logic in HumanInTheLoopMiddleware

[Saakshi Gupta (saakshigupta2002)]

Summary

This PR fixes a RuntimeError that occurs when using HumanInTheLoopMiddleware with async agent invocation (.ainvoke()). The middleware's aafter_model method was delegating directly to the synchronous after_model method, which broke the async runnable context chain required by langgraph's interrupt() function.

Fixes #34974

Problem

When an agent using HumanInTheLoopMiddleware is invoked via .ainvoke(), the following error is raised:

RuntimeError: Called get_config outside of a runnable context

Stack trace:

File ".../human_in_the_loop.py", line 381, in aafter_model
    return self.after_model(state, runtime)
File ".../human_in_the_loop.py", line 331, in after_model
    decisions = interrupt(hitl_request)["decisions"]
File ".../langgraph/types.py", line 515, in interrupt
    conf = get_config()["configurable"]
File ".../langgraph/config.py", line 29, in get_config
    raise RuntimeError("Called get_config outside of a runnable context")

Reproduction steps:

from langchain.agents.middleware.human_in_the_loop import HumanInTheLoopMiddleware
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver

middleware = HumanInTheLoopMiddleware(interrupt_on={"my_tool": True})
agent = create_react_agent(
    model, tools, middleware=[middleware], checkpointer=MemorySaver()
)

# This works fine (sync):
agent.invoke({"messages": [("user", "use my_tool")]}, config=config)

# This fails with RuntimeError (async):
await agent.ainvoke({"messages": [("user", "use my_tool")]}, config=config)

Root Cause

The aafter_model async method was implemented as a simple delegation to the sync after_model:

async def aafter_model(self, state, runtime):
    return self.after_model(state, runtime)  # Bug: loses async context

When langgraph's async executor calls aafter_model, it sets up an async runnable context (via contextvars). The interrupt() function, called inside after_model, relies on get_config() to access this context. However, since after_model is called as a plain synchronous function from within the async coroutine, the async runnable context is not properly accessible, causing the RuntimeError.

This issue is particularly pronounced on Python 3.10 where contextvars propagation in asyncio differs from later versions.

Solution

Inlined the full after_model logic directly into aafter_model instead of delegating to the sync method. The logic is identical -- no async-specific changes are needed since interrupt() is synchronous. The key fix is ensuring that interrupt() -> get_config() executes directly within the aafter_model coroutine body, where the async runnable context is accessible.

Changes

libs/langchain_v1/langchain/agents/middleware/human_in_the_loop.py

• Replaced the one-line delegation in aafter_model with the full inlined logic from after_model
• The method now independently handles: message inspection, action request/config creation, interrupt flow, decision validation, and tool call reconstruction

libs/langchain_v1/tests/unit_tests/agents/middleware/implementations/test_human_in_the_loop.py

Added 6 new async unit tests for the aafter_model path:

Test	Description
test_aafter_model_no_interrupts_needed	Verifies early returns for empty messages, no tool calls, and non-matching tools
test_aafter_model_single_tool_accept	Tests approval flow through async path
test_aafter_model_single_tool_edit	Tests edit flow with modified args through async path
test_aafter_model_single_tool_reject	Tests rejection flow with tool message generation
test_aafter_model_multiple_tools_mixed_responses	Tests mixed approve/reject across multiple tools
test_aafter_model_preserves_tool_call_order	Verifies original tool call order is preserved with interleaved auto-approved and interrupted tools

Test Plan

• All 18 existing sync tests continue to pass
• All 6 new async tests pass
• Full test suite: 24/24 passed

How to verify the fix end-to-end

import asyncio
from langchain.agents.middleware.human_in_the_loop import HumanInTheLoopMiddleware
from langgraph.prebuilt import create_react_agent
from langgraph.checkpoint.memory import MemorySaver

middleware = HumanInTheLoopMiddleware(interrupt_on={"my_tool": True})
agent = create_react_agent(
    model, tools, middleware=[middleware], checkpointer=MemorySaver()
)

# Previously failed with RuntimeError, now works correctly:
result = asyncio.run(agent.ainvoke(
    {"messages": [("user", "use my_tool")]},
    config={"configurable": {"thread_id": "1"}}
))