HITL: Resuming a pipeline with an agent using the BreakpointConfirmationStrategy fails

[Hansehart]

Summary

When using Haystack Experimental HITL with standalone agents, resuming agent execution fails with:

KeyError: 'serialized_data'

This happens because haystack_experimental creates snapshots with:

original_input_data = {}

But core Haystack expects the snapshot to include:

{
  "serialization_schema": {...},
  "serialized_data": {...}
}

Steps to Reproduce

1. Run an agent with HITL/tool approval enabled (using haystack-experimental).

2. The agent triggers a breakpoint and returns an approval snapshot.

3. Submit the approval decision and attempt to resume.

4. Resume fails with KeyError: 'serialized_data'.

Solution

In haystack_experimental/core/pipeline/breakpoint.py, snapshots for standalone agents are created with:

original_input_data={}

However it should be:

original_input_data=_serialize_value_with_schema({}),

[sjrl]

Hey @Hansehart could you provide a code example to reproduce this error? We have an integration test here

haystack-experimental/test/components/agents/test_agent.py

Line 307 in 0b4f5e0

async def test_run_async_breakpoint_confirmation_strategy_modify(self, tools, tmp_path):

that runs as expected using a standalone agent.

[Hansehart]

@sjrl I cant reproduce an example fast. However could you confirm if this analysis is correct?

Looking at the source code, I believe the issue might be in haystack_experimental/core/pipeline/breakpoint.py in the
_create_pipeline_snapshot_from_tool_invoker function:

# Line 114 - When parent_snapshot is None (standalone agents):
if parent_snapshot is None:
    final_snapshot = PipelineSnapshot(
        ...
        original_input_data={},  # ← Is this the issue? Empty dict without serialization structure?
        ...
    )

When Haystack tries to resume from the snapshot, it seems to expect a specific structure (https://github.com/deepset-ai/haystack/blob/main/haystack/core/pipeline/breakpoint.py#L88):

serialized_input_data = pipeline_snapshot.original_input_data["serialized_data"]
# ↑ This would throw KeyError because original_input_data is just {}

However could be an layer 8 failure too :) it was just an experimental setup to try out the hitl logic. Maybe I run into some edge cases.

[sjrl]

@Hansehart thanks for the additional info! I still don't see how this would occur since the code https://github.com/deepset-ai/haystack/blob/main/haystack/core/pipeline/breakpoint.py#L88 is only called when restarting a pipeline with a pipeline snapshot.

And the code path in _create_pipeline_snapshot_from_tool_invoker that looks problematic should only occur if you are running the Agent standalone as a component and not in a pipeline.

I agree that we should make these consistent so either update original_input_data={} or the check in _validate_break_point_against_pipeline but currently I don't see how this would be triggered.

[Hansehart]

I tried to recreate my sandbox and created a simple agent in a pipeline:

from haystack import Pipeline
from haystack.tools import Tool
from haystack.dataclasses import ChatMessage
from haystack_experimental.components.agents import Agent
from haystack_experimental.components.agents.human_in_the_loop import (
    BreakpointConfirmationStrategy,
)
from haystack.core.errors import BreakpointException
from haystack.core.pipeline.breakpoint import load_pipeline_snapshot
from haystack_integrations.components.generators.mistral import MistralChatGenerator


# Define a regular function
def greet(name: str) -> str:
    """Greet someone by name."""
    return f"Hello {name}"


# Simple tool with HITL
tool = Tool(
    name="greet",
    description="Greet someone",
    function=greet,
    parameters={
        "type": "object",
        "properties": {"name": {"type": "string"}},
        "required": ["name"],
    },
)

# Agent with HITL inside Pipeline
agent = Agent(
    chat_generator=MistralChatGenerator(model="ministral-3b-2512"),
    tools=[tool],
    confirmation_strategies={
        "greet": BreakpointConfirmationStrategy(snapshot_file_path="snapshots")
    },
)

pipeline = Pipeline()
pipeline.add_component("agent", agent)

# Step 1: First run triggers HITL
try:
    pipeline.run({"agent": {"messages": [ChatMessage.from_user("Greet Alice")]}})
except BreakpointException as e:
    print(f"HITL triggered, snapshot: {e.pipeline_snapshot_file_path}")

    # Step 2: Try to resume -> KeyError: 'serialized_data'
    snapshot = load_pipeline_snapshot(e.pipeline_snapshot_file_path)
    pipeline.run({"agent": {"messages": []}}, pipeline_snapshot=snapshot)

Even when an Agent runs inside a Pipeline, on the first run it doesn't receive a parent_snapshot. This triggers the bug. Wdyk @sjrl ?

[sjrl]

Once deepset-ai/haystack#10222 is merged and released in a new version of Haystack we will fix this issue in haystack-experimental