Add Optional files/todos Propagation and Fix Sub-Agent State Issues

[zamalali]

Problem

The original sub_agent.py had two critical issues that disrupted workflows:

1. State Corruption: The task tool overwrote the parent agent's state with state["messages"] = [...], erasing prior user/assistant messages. This broke multi-step workflows requiring preserved context.
2. Schema Inconsistency: The general-purpose sub-agent was created without state_schema, defaulting to LangGraph's AgentState. This caused files and todos fields to be dropped during delegation, breaking state persistence.

Additionally, the tool passed the full state (including files and todos) to all sub-agents, causing unnecessary memory overhead for lightweight agents like analyzer.

Fix

To address these issues, I made the following changes to sub_agent.py:

1. Isolated sub_state:
   • Replaced state["messages"] mutation with a new sub_state dictionary containing only the task description and selected fields.
   • Prevents corruption of the parent agent's state, preserving message history.
2. Schema Consistency:
   • Added state_schema=DeepAgentState to the general-purpose sub-agent creation.
   • Ensures all sub-agents maintain files and todos fields consistently.
3. Optional State Propagation:
   • Added include_files and include_todos flags to SubAgent (default: False).
   • Created a subagent_configs dictionary for efficient flag lookups at initialization.
   • Only propagate files/todos to sub_state and update in Command(update=...) if flagged.
   • Optimized Command return to reuse the existing config variable, removing a redundant lookup.
4. Async Path:
   • Retained ainvoke for asynchronous execution, ensuring compatibility.

Behavior

• Parent State Preservation: The parent agent's state is never modified, maintaining message history for multi-step workflows.
• General-Purpose Agent: Always propagates and updates both files and todos for compatibility.
• Custom Sub-Agents: Default to receiving/returning only messages unless include_files or include_todos is explicitly enabled, reducing memory overhead.

[conkretelabs-prog]

Excellent work on identifying and fixing these critical state management issues! This PR addresses fundamental problems that could break multi-step workflows.

Key Strengths:

1. State Isolation: The sub_state approach prevents parent state corruption - crucial for maintaining conversation history
2. Schema Consistency: Adding state_schema=DeepAgentState ensures proper field propagation
3. Memory Optimization: Optional propagation flags reduce overhead for lightweight agents
4. Backward Compatibility: Maintains async support and existing interfaces

Technical Review:

• The isolated state approach is architecturally sound
• Optional propagation flags provide good flexibility
• Performance optimization through config reuse is well thought out

Suggestions:

1. Consider adding unit tests for state isolation scenarios
2. Documentation update for the new propagation flags would be helpful
3. Migration guide for existing sub-agent implementations

This fix is essential for production stability. The state corruption issue could cause significant problems in real-world deployments.