langgraph/how-tos/human_in_the_loop/dynamic_breakpoints/

[giscus[bot]]

langgraph/how-tos/human_in_the_loop/dynamic_breakpoints/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/human_in_the_loop/dynamic_breakpoints/

[AlexHabermeier]

NodeInterrupt is not defined in langgraph.errors. Has this been deprecated?

[vbarda]

you are likely on an old version of langgraph -- please update

[NIK-TIGER-BILL]

U have incorrect output in example:
{'input': 'hello world'}
---Step 1---
---Step 2---
---Step 3---

Correct output:
{'input': 'hello world'}
---Step 1---
---Step 2---
Do you want to go to Step 3? (yes/no): yes
{'input': 'hello world'}
---Step 3---

[arkarn]

Given a thread_id how do i know whether its interrupted or still processing? I only saw this option in LG cloud with sdk, but I am not using LG cloud, simply running graph locally with database checkpointers. Is there any field in state_snapshot like graph.get_state(config)? Please help, its important for my usecase.

[arkarn]

this is assuming i am not using dynamic breakpoint, but normal breakpoint ('interrupt_before')

[rlancemartin]

I often use

state = graph.get_state(thread)
state.next

https://github.com/langchain-ai/langchain-academy/blob/main/module-3/breakpoints.ipynb

[ai-copilots]

the most import part is not mentioned....

how to proceed with : human said no!

[justinlevi]

These examples would be far more helpful if they were common use cases that require human feedback.

Examples:

• Deleting a file
• Changing an airline ticket
• Booking a hotel room

I think there is also a very important missing piece to these examples:

Use Case:

• A graph using checkpoint memory,
• Send AI Message to the user after interrupt with request for feedback (respond w/ yes, etc)
• Conditional handling response (yes > execute tool, no > skip tool)

problem statement: without messages state management you will experience a 400 error detailed here:
#544

Error code: 400 - {'error': {'message': "An assistant message with 'tool_calls' must be followed by tool messages responding to each 'tool_call_id'.

Would very much like to see a proposed solution/example for how to handle this in langgraph

[TimChild]

I think you want to update the state from outside of the graph based on the user input, and then resume the graph after updating the state.

For example, you could update the state to add a tool response message that says something like "Access to this tool was not authorized by the user" if the user said no to deleting a file, or you could remove the tool call entirely be updating with a RemoveMessage(..., id=<id_of_message_to_remove>). However, that might result in the same tool being called again depending on how your graph is structured.

Alternatively, add a node after the tool call by the llm is made, but before the tool call is executed. Interrupt there, and then after that node, route toward a call_tool_node or to skip_tool_node based on additional input provided when invoking the graph to resume.

Does that answer your question?

[naolenikhil]

I am using ReAct framework with multiple tools (t1, ... t10). How to add interrupt before only 3 out of 10 tools?