Lár v1.4.1 Release Notes

Lár v1.4.1 Release Notes

(Patch Release - February 15, 2026)

This patch release addresses a critical bug in audit logging and formalizes support for Reasoning Models (System 2 thinking).

Critical Fixes

1. Robust Audit Logging (Fixes "Missing Logs")

Problem: In v1.4.0, if a script exited early (e.g., user break in a loop, sys.exit, or crash), the GraphExecutor would bypass the log-saving step. Fix: Wrapped the execution loop in a try...finally block. Impact: Logs are now always saved to lar_logs/run_{uuid}.json (or your custom log_dir), guaranteeing a complete audit trail.

New Features

1. Reasoning Model Support (System 2 Thinking)

Lár now automatically detects and captures "Reasoning Traces" (Chain of Thought), keeping your final answer clean while preserving the thought process for auditing.

Supported Models:

• DeepSeek R1 (API & Distilled/Ollama)
• OpenAI o1 (Preview/Mini)
• Liquid Thinking (ollama/liquid-thinking)

How it Works:

1. Metadata Capture: If the API returns reasoning_content (Standard), it saves to state["__last_run_metadata"]["reasoning_content"].
2. Robust Regex Fallback: If the model outputs raw <think>...</think> tags (common in local models), Lár extracts them.
   • Robustness: Handles missing closing tags (cut-off thoughts) and missing opening tags (hallucinated starts).
   • Clean State: The main output_key (e.g., state['answer']) contains only the final response. The reasoning is safely stored in metadata.

2. New Examples

Added a dedicated directory examples/reasoning_models/ with patterns for:

• 1_deepseek_r1.py: DeepSeek R1 / Generic Ollama.
• 2_openai_o1.py: OpenAI o1 Series.
• 3_liquid_thinking.py: Liquid Thinking (Lateral Logic).

Internal Changes

• executor.py: Added finally block for log persistence.
• node.py: Enhanced LLMNode with robust regex for tag extraction.
• tracker.py: valid logic confirmed for per-model token tracking.

Lár v1.4.1 is a recommended update. It ensures your logs are safe and your reasoning models work out-of-the-box.