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

Overview

Simon Willison's open-source command-line tool LLM has released version 0.32a1, a quick hotfix for version 0.32a0 that resolves a data integrity issue when restoring tool-calling sessions from the SQLite database.

The Fix: SQLite Storage Issue for Tool-Calling Sessions

The core fix in this update addresses Issue #1426: in version 0.32a0, when users engaged in conversations using the tool-calling feature, session records could be written to the SQLite database without issue, but errors occurred when reinflating them from the database, preventing historical sessions from being correctly restored.

This issue significantly impacts users who rely on session persistence. One of LLM's core selling points is storing all conversation records in a local SQLite database, making it easy for users to review, search, and continue conversations later. As an important new feature introduced in version 0.32, the reliability of tool-calling session storage directly affects the day-to-day user experience.

The Complexity of Session Serialization and Deserialization

Storing tool-calling sessions is far more complex than storing ordinary conversations. A typical conversation has a simple "user message → assistant reply" alternating structure, while a tool-calling session's message sequence may include: user messages, the assistant's tool call requests (structured data containing function names and parameters), tool execution results (tool results), and the assistant's final reply generated based on those tool results. These different message types need to retain complete type information and nested structures when serialized to the database, and each message's role and content format must be accurately restored during deserialization (i.e., reinflation). Any missing fields or type errors will cause session restoration to fail—which is exactly the problem exposed by Issue #1426.

The Design Philosophy of SQLite as Local Data Storage

Simon Willison is a long-time advocate for SQLite, which he considers one of the "most underrated technologies." Unlike traditional client-server databases (such as PostgreSQL or MySQL), SQLite is an embedded database engine where the entire database is a single file that requires no separate server process. This makes it ideal as a local storage solution for command-line tools—users don't need to configure any database service, and all conversation records are saved in a single portable .db file. Simon Willison also developed the Datasette project, specifically designed for exploring and publishing data from SQLite databases. LLM's storage design is fully aligned with his overall technical philosophy.

Introduction to the LLM CLI Tool

LLM is a command-line tool built by well-known developer Simon Willison that lets users interact with various large language models directly from the terminal. It supports multiple model providers including OpenAI, Anthropic, and Google, and achieves high extensibility through a plugin system.

Core Features

• Multi-model support: Supports virtually all major LLM providers through a plugin mechanism
• Session management: All conversations are automatically stored in a local SQLite database
• Tool calling: Supports function calling / tool use, enabling models to invoke external tools to complete tasks
• Template system: Supports custom prompt templates to boost daily workflow efficiency

Plugin Ecosystem

LLM's plugin system is built on Python's pluggy framework, the same plugin architecture used by well-known projects like pytest. The community has already developed dozens of plugins covering Anthropic Claude (llm-claude-3), Google Gemini (llm-gemini), locally-run Ollama models (llm-ollama), Mistral, Groq, and nearly all other major model services. Users can simply run llm install llm-claude-3 to add support for a new model without modifying the core code. This design allows LLM to quickly adapt to the rapidly evolving pace of model releases in the AI field, and community contributors can independently maintain their plugins without affecting the main project's stability.

Why Tool Calling Matters

Tool calling is one of the core capabilities in current LLM application development. It allows models to identify scenarios during conversations that require external data or actions, and automatically invoke predefined functions to retrieve information or execute tasks. LLM CLI introduced this feature in version 0.32, enabling command-line users to experience Agent-style interactions—such as having the model query databases, read files, or call APIs.

Technical Principles of Tool Calling

Tool calling (also known as Function Calling) has been one of the most important capability advancements in the LLM API space since 2023. Its core principle is: when generating a response, the model can output not only natural language text but also structured function call requests (typically in JSON format) containing function names and parameters. The application layer receives this request, executes the corresponding function, and returns the result to the model, which then continues generating the final response based on the function's return value. This process forms a "model → tool → model" loop and is the foundational architecture for building AI Agents. OpenAI pioneered the Function Calling API in June 2023, followed by Anthropic's Tool Use, Google's Function Calling, and others—it has now become an industry-standard capability.

Version Notes and Considerations

0.32a1 is still an alpha version (the "a" in the version number stands for alpha), meaning the feature is still under active development and testing. If you're using LLM in a production environment, it's recommended to watch for subsequent stable releases.

Alpha Versions and Python Version Control

The Python ecosystem follows the PEP 440 versioning specification, where 'a' denotes an alpha version, 'b' denotes a beta version, and 'rc' denotes a release candidate. An alpha version means the feature is mostly implemented but may have known issues, primarily targeting early testers and contributors. In pip's default behavior, alpha and beta versions are not automatically installed—running pip install llm will only install the latest stable version, and you must explicitly specify a version number or use the --pre flag to install pre-release versions. This mechanism protects regular users from accidentally upgrading to an unstable version while allowing those who want to try new features to do so early and provide feedback.

This quick fix also demonstrates Simon Willison's attitude toward project quality—rapidly releasing a patch after discovering an issue to ensure early testers get a stable experience.

How to Install or Update

Use pip to install or update to the latest alpha version:

pip install llm==0.32a1

If you want to always track the latest pre-release version:

pip install --pre llm

Key Takeaways

• LLM CLI released version 0.32a1 alpha, fixing a critical bug from 0.32a0
• The fixed issue involves incorrect data when restoring tool-calling sessions from the SQLite database (Issue #1426); the root cause is that tool-calling session message structures are more complex than ordinary conversations, and data was lost during serialization and deserialization
• Tool calling is a major new feature introduced in version 0.32, based on the industry-standard Function Calling protocol, enabling models to invoke external tools to complete tasks
• LLM uses SQLite as its local storage solution, reflecting Simon Willison's consistent "SQLite-first" design philosophy
• This version is still in alpha stage; pip won't install pre-release versions by default, and production environments should wait for the stable release