Lár v1.3.0 Release Notes

Lár v1.3.0 Release Notes

"The Compliance & Architecture Update"

This release focuses on hardening the framework for enterprise and regulatory compliance (EU AI Act), while refactoring the core execution engine for better separation of concerns.

Key Features

1. "Human-in-the-Loop" Primitive (HumanJuryNode)

A new node type that pauses execution to request explicit human feedback via the CLI.

• Why: Directly satisfies EU AI Act Article 14 ("Human Oversight") requirements.
• How:

  jury = HumanJuryNode(
      prompt="Approve sensitive action?",
      choices=["approve", "reject"],
      output_key="approval_status"
  )
  

2. Static Safety Analysis (TopologyValidator)

We've added a TopologyValidator (backed by NetworkX) that runs comprehensive checks on Dynamic Graphs before they execute.

• Cycle Detection: Catches infinite loops in generated subgraphs.
• Structural Integrity: Validates next_node references.
• Tool Allowlisting: Enforces strict boundaries on what tools a dynamic agent can access.

3. Core Refactor: Modular Observability

The GraphExecutor has been refactored to delegate responsibilities to dedicated components, keeping the engine lightweight.

• AuditLogger: Centralizes audit trail logging and file persistence.
• TokenTracker: Aggregates token usage across multiple providers and models with precision.

Usage Updates

Breaking Changes

• GraphExecutor constructor now accepts optional logger and tracker instances for dependency injection.
• Compliance: The "Glass Box" is now even more transparent with improved metadata fidelity in logs.

How to Use (New in v1.3.0)

Option 1: Automatic (Default Behavior)

from lar import GraphExecutor, LLMNode

# Logger and Tracker are created automatically
executor = GraphExecutor(log_dir="my_logs")

node = LLMNode(model_name="ollama/phi4", prompt_template="test", output_key="result")
result = executor.run(node, {})

# Access automatically created instances
print(executor.logger.get_history())  # Audit trail
print(executor.tracker.get_summary())  # Token usage

Option 2: Custom Injection (Advanced)

from lar import GraphExecutor, AuditLogger, TokenTracker

# Create custom instances
custom_logger = AuditLogger(log_dir="advanced_logs")
custom_tracker = TokenTracker()

# Inject into executor
executor = GraphExecutor(
    logger=custom_logger,
    tracker=custom_tracker
)

# Share tracker across multiple executors for aggregated cost tracking
executor2 = GraphExecutor(
    logger=AuditLogger(log_dir="other_logs"),
    tracker=custom_tracker  # Same tracker = aggregated tokens
)

Why Custom Injection?

• Centralized audit trail management
• Cost aggregation across workflows
• Custom log formatting/persistence
• Integration with existing monitoring systems

See: examples/patterns/16_custom_logger_tracker.py for full demo

Chains to State Machines

With v1.3, the debate is settled. Lár's state machine architecture now offers native compliance features (Breakpoints, Static Analysis) that Chain-based frameworks struggle to implement.

Changelog

• [NEW] HumanJuryNode in src/lar/node.py
• [NEW] TopologyValidator in src/lar/dynamic.py
• [NEW] AuditLogger in src/lar/logger.py
• [NEW] TokenTracker in src/lar/tracker.py
• [REFACTOR] GraphExecutor to use new helper classes.