[DEMO - SCALE 23X] Modifiable node inputs

[mkmeral]

Problem Statement

Please note that this is a temporary ticket meant for live demo during Scale 23X conference

Allow to customize node inputs for graph structure

Proposed Solution

No response

Use Case

I want to be able to customize my node inputs in graphs, before they are sent to the agent

Alternatives Solutions

No response

Additional Context

No response

[mkmeral]

/strands

[github-actions]

Repository Analysis & Clarifying Questions

I've analyzed the repository structure and the graph implementation. Here's what I found and some questions to ensure proper implementation:

Repository Context

• Location: src/strands/multiagent/graph.py
• Key Method: _build_node_input() (lines 1030-1113) - automatically builds node input from dependency outputs or original task
• Current Hook: BeforeNodeCallEvent fires before each node executes, but does NOT expose the node input for modification
• Input Types: list[ContentBlock] (from src/strands/types/content.py)

Proposed Implementation Approach

Extend BeforeNodeCallEvent to include a modifiable node_input field:

1. Build node input before firing BeforeNodeCallEvent
2. Include node_input in the event (writable field)
3. Use the (potentially modified) input from the event when executing the node

Example usage:

from strands.hooks import BeforeNodeCallEvent, HookProvider

class InputModifierHook(HookProvider):
    def register_hooks(self, registry, **kwargs):
        registry.add_callback(BeforeNodeCallEvent, self.modify_input)
    
    def modify_input(self, event: BeforeNodeCallEvent):
        # Access and modify the node input before agent execution
        if event.node_id == "target_node":
            event.node_input = [ContentBlock(text="Custom input for this node")]
        return event

Clarifying Questions

Functionality

1. Scope of modification: Should this feature allow modifying inputs for ALL nodes, or should there be an option to restrict which nodes can have their inputs modified?

2. Original input preservation: Should the original (auto-built) input also be accessible in the event alongside the modifiable copy? (e.g., original_node_input as read-only, node_input as modifiable)

3. Input validation: Should the modified input be validated (e.g., type checking for list[ContentBlock]) before being passed to the node executor?

Integration

4. Interaction with invocation_state: Should the modified node input integrate with the existing invocation_state mechanism, or remain separate?

5. Interrupt resumption: When resuming from an interrupt, should the hook also fire with the interrupt response as node_input? (Currently, interrupt responses bypass the normal input building logic)

Performance

6. Lazy vs eager evaluation: Building node input requires accessing dependency results. Is it acceptable to always build the input before firing the hook, even if it won't be modified? (This is the simpler approach)

---

Estimated Scope: This appears to be a small-to-medium sized feature that can be implemented in a single PR. Files to modify:

• src/strands/hooks/events.py - Add node_input field to BeforeNodeCallEvent
• src/strands/multiagent/graph.py - Pass and use modified node_input
• Tests in tests/strands/multiagent/test_graph.py and tests/strands/hooks/test_events.py

[mkmeral]

/strands implement using whatever you prefer, i trust you

[github-actions]

Unable to create pull request via API. You can create it manually by clicking here.

[mkmeral]

This was a test. Not bad, but could be better 😅