LLM CLI Tool 0.32a1 Released: Fixes Tool-Calling Conversation Storage Bug

Overview

Simon Willison has released version 0.32a1 (preview) of his open-source command-line tool LLM, primarily fixing a bug from the previous version (0.32a0) where tool-calling conversations failed to restore correctly from the SQLite database.

Core Fix: SQLite Storage Issue for Tool-Calling Conversations

This update addresses Issue #1426. The specific problem was: after users conducted conversations using LLM's tool-calling feature, the conversation records would be persisted to a local SQLite database. However, in version 0.32a0, when reinflating these conversations containing tool-calling information from the database, the data could not be correctly restored, causing conversation history anomalies.

This fix is particularly critical for developers who rely on tool-calling functionality. Function calling is one of the core capabilities in modern LLM applications, allowing models to invoke external functions or APIs during conversations. The ability to fully record and trace back conversations is fundamental for debugging and reuse.

Technical Background on Tool Calling

Tool calling is a key technology in the evolution of large language models from pure text generation to actionable capabilities. The core principle is: when generating a response, the model can output structured function call requests (including function names and parameters), which are executed by an external system that returns results to the model, enabling it to continue generating responses based on execution results. OpenAI first introduced function calling in GPT models in June 2023, followed by Anthropic, Google, and other providers. This capability enables LLMs to query databases, call APIs, manipulate file systems, and more—making it foundational infrastructure for building AI Agents.

Why SQLite Storage of Tool-Calling Conversations Is Error-Prone

SQLite is an embedded relational database that requires no separate server process, storing the entire database in a single file. Simon Willison is a long-time advocate of SQLite (he's also the creator of Datasette), and choosing SQLite as the conversation storage backend for the LLM tool ensures both a zero-configuration out-of-the-box experience and structured query capabilities. However, the data structure of tool-calling conversations is far more complex than plain text conversations—it includes function call requests, parameter JSON, execution results, and other nested information. The serialization and deserialization (reinflation) process is prone to edge cases, which was the root cause of the bug in version 0.32a0.

Introduction to the LLM Tool

LLM is a command-line tool and Python library developed by Simon Willison that allows users to interact with multiple large language models through a unified interface. It supports various backends including OpenAI, Anthropic, and local models, with high extensibility through a plugin mechanism. The recently added tool-calling support has significantly enhanced its utility in AI Agent and automation workflow scenarios.

In the context of the explosive growth of the AI Agent ecosystem between 2024-2025—from AutoGPT and LangChain to CrewAI and the OpenAI Agents SDK, with frameworks emerging continuously—command-line LLM interaction tools play a unique role: they serve as powerful instruments for developers' daily debugging and rapid prototyping, while also seamlessly integrating into automation workflows through shell scripts and piping mechanisms. LLM's plugin architecture enables it to adapt to different model providers, and tool-calling support elevates it from a simple Q&A tool to an orchestrable Agent component.

Version Notes and Installation

It's important to note that 0.32a1 is still an Alpha preview version (indicated by the "a" suffix), primarily targeting early testers. For production environments, it's recommended to watch for subsequent stable releases.

The Python ecosystem follows the PEP 440 versioning specification, where the "a" suffix denotes an Alpha version, "b" denotes Beta, and "rc" denotes Release Candidate. An Alpha version means features may be incomplete or have known issues, and is mainly used to gather early feedback. pip does not install pre-release versions by default—users must explicitly specify the version number or use the --pre flag to obtain them. This release strategy allows developers to iterate quickly and fix issues without affecting stable users.

To install or upgrade to this version:

pip install llm==0.32a1

Or use the --pre flag to get the latest pre-release version:

pip install --pre llm

Summary

Although this is a minor version update, it reflects the ongoing refinement of the LLM tool's function calling capabilities. As the AI Agent ecosystem rapidly evolves, command-line LLM interaction tools are becoming an indispensable part of the developer toolchain. For engineers focused on AI development efficiency, the LLM tool is worth following closely.

Key Takeaways

• LLM 0.32a1 fixes a bug in version 0.32a0 where tool-calling conversations were incorrectly restored from the SQLite database
• The issue affected the persistence and retrieval of conversation histories containing tool-calling information
• Tool-calling conversation data structures include function names, parameter JSON, execution results, and other nested information, making serialization far more complex than plain text conversations
• LLM is an open-source command-line tool and Python library supporting multiple model backends, with extensibility through a plugin mechanism
• The current version is an Alpha preview—pip won't install it by default, requiring explicit version specification