LLM 0.32 发布：新增推理轨迹、OpenAI Responses、服务端工具与更智能的日志

我今天早上发布了 LLM 0.32：https://llm.datasette.io/en/stable/changelog.html#v0-32，这是 LLM 自项目初次发布以来最重要的新版本。新版本包括对可见推理轨迹的支持、服务器端提供者工具、重新设计的内容寻址 SQLite 日志、新模型以及通过 OpenAI Responses API 启用的新功能。我还发布了 llm-anthropic 插件的新版本：https://github.com/simonw/llm-anthropic，自身也有大量更新。

在推理模型上运行 LLM 现在会将它们的推理轨迹显示到标准错误输出，因此你可以看到它们的“思考”内容，而这些信息不会包含在你可能传给其他工具的标准输出中。使用 -R/--hide-reasoning 可以关闭此功能。

LLM 开箱即用地支持 GPT-5.6 模型系列，并且使用 llm "prompt" 的新默认模型现在是廉价但功能强大的 GPT-5.6 Luna。

LLM 调用现在可以使用来自各种提供者的服务器端工具。OpenAI 提供了一个代码执行环境：https://llm.datasette.io/en/stable/openai-models.html#code-interpreter 作为服务器端工具；LLM 现在可以运行利用该工具受益的提示，如下所示：

OpenAI 还获得了一个 WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search 工具。

llm-anthropic：https://github.com/simonw/llm-anthropic 插件增加了 WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search、WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch、CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution，以及 AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector，看起来是这样的：

这样会导致 Anthropic 在与其 API 的单次请求/响应交互中，对我的新 datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp 插件执行 MCP 调用。

新的 llm openai endpoint 命令提供了一个工具，可以对任何兼容 OpenAI 的端点执行提示：https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it，一行命令即可完成。这些操作不会被记录，因此它是一个便捷的工具，可以对任何使用 LLM API 世界通用语言的对象运行一次性提示。

这是我如何使用它，通过 uvx（无需安装 LLM）并结合 llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs 工具插件，对运行在我本地主机 LM Studio：https://lmstudio.ai 的 Gemma 4 12B 进行提示的方式：

LLM 的 Python API 以前要求你创建一个会话，然后一次发送一条消息。这是对 LLM 真正特性的抽象，每个请求都携带之前所有消息的完整历史。对于一些更高级的情况，这种抽象开始成为障碍，因此新版本引入了 model.prompt(messages=[]) 参数，可以这样使用：

LLM 以前会从每个提示返回一个字符串的可迭代序列。当模型返回字符串响应时，这效果很好，但无法预测模型会演变出的奇怪形式。如今，许多模型返回推理文本、输出字符串、工具调用，甚至图像附件的混合内容。在 LLM 0.32 中，你可以改用这个方法：https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events

结合这些功能，我们终于可以提供一个稳健的半标准 OpenAI 聊天完成 API 实现，我已经将其作为 llm-chat-completions-server：https://github.com/simonw/llm-chat-completions-server 插件发布了：

现在，你可以通过该服务器使用新的 llm openai endpoint 命令来对 LLM 运行提示！

这种类型的 API 更大的挑战在于日志记录。如果我们要支持每次请求都追加消息序列的模式，理想情况下可以避免为每次会话记录所有重复的 JSON。

解决方案是新的内容可寻址消息存储：https://llm.datasette.io/en/stable/logging.html#the-message-store， 模仿 Git 建模。你可以在文档中看到新的模式：https://llm.datasette.io/en/stable/logging.html#sql-schema，但 llm logs 和 llm logs --json 命令都已升级，可以将该格式转换回易于使用的形式。

本次发布还有更多内容。0.32 发布说明：https://llm.datasette.io/en/stable/changelog.html#v0-32 相当全面，0.32rc2：https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30、0.32rc：https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30、0.32a3：https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09、0.32a2：https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12 和 0.32a0：https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 的说明可以填补任何遗漏。

现有的 LLM 插件应该都可以继续使用，但提供额外模型的插件需要升级到 0.32 才能完全参与新的流事件系统。文档中有关于使用结构化消息和流事件实现插件的指南：https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events。

我更新了一些我自己的插件：

本次发布的许多底层工具更改是由 Datasette Agent：https://agent.datasette.io/ 的需求驱动的。当我开始开发 LLM 时，“代理”这个术语定义非常模糊，我当时拒绝使用它。到 2025 年 9 月：https://simonwillison.net/2025/Sep/18/agents/ 我接受了这样一个观点：“LLM 代理通过循环运行工具以实现目标” 已经足够成熟，我可以完全停止回避这个术语。

工具链现在可以暂停以等待人工批准：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause，并从存储的消息历史中恢复：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume——这两点都是 Datasette Agent 所需的。

今天看LLM，它开始在我看来呈现出非常像代理（agent）的形态。有一个很妙的地方是，你可以用一个命令行工具将来自不同来源、使用不同模型的各种工具混合搭配，一行代码就能搞定，而且它包括一个足够强大的Python库，可以构建像Datasette Agent（https://agent.datasette.io/）和llm-coding-agent（https://github.com/simonw/llm-coding-agent）这样的系统。

也许下一版本的LLM会将“代理”的概念直接融入核心库。我仍在尝试弄清楚那会是什么样子。

这是LLM的新版本，由Simon Willison在2026年8月4日发布：新增对推理追踪、OpenAI响应、服务器端工具以及更智能日志记录的支持：/2026/Aug/4/。

系列文章“大语言模型（LLM）新版本发布”的一部分：/series/llm-releases/

上一篇：无状态MCP重新引起了我的兴趣（并启发了mcp-explorer和datasette-mcp）：/2026/Jul/31/stateless-mcp/

每月赞助我10美元，即可收到精选邮件摘要，汇总当月最重要的LLM动态。

I released LLM 0.32：https://llm.datasette.io/en/stable/changelog.html#v0-32 this morning, the most significant new version of LLM since the initial launch of the project. The new version includes support for visible reasoning traces, server-side provider tools, redesigned content-addressable SQLite logs, new models, and new features enabled by the OpenAI Responses API. I also released a new version of the llm-anthropic plugin：https://github.com/simonw/llm-anthropic with substantial updates of its own.

Running LLM against reasoning models now displays their reasoning traces to standard error, so you can see what they are “thinking” without that information being included in the standard output that you might pipe to another tool. Add -R/--hide-reasoning to turn this off.

LLM includes support out-of-the-box for the GPT-5.6 model family , and the new default model used with llm "prompt" is now the inexpensive but capable GPT-5.6 Luna .

LLM calls can now use server-side tools from various providers. OpenAI provide a code execution environment：https://llm.datasette.io/en/stable/openai-models.html#code-interpreter as a server-side tool; LLM can now run prompts that benefit from that like so:

OpenAI also gets a WebSearch：https://llm.datasette.io/en/stable/openai-models.html#web-search tool.

The llm-anthropic：https://github.com/simonw/llm-anthropic plugin adds WebSearch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-search, WebFetch：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#web-fetch, CodeExecution：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#code-execution, and AnthropicMCP：https://github.com/simonw/llm-anthropic/blob/0.26/README.md#mcp-connector, which looks like this:

That causes Anthropic to execute MCP calls against my new datasette-mcp：https://simonwillison.net/2026/Jul/31/stateless-mcp/#datasette-mcp plugin as part of a single request/response interaction with their API.

The new llm openai endpoint command provides a tool for executing prompts against any OpenAI compatible endpoint：https://llm.datasette.io/en/stable/other-models.html#run-against-an-endpoint-without-configuring-it as a one-liner. These aren’t logged, which makes this a handy tool for running one-off prompts against anything that speaks the lingua franca of the LLM API world.

Here’s how I use that to run prompts against Gemma 4 12B running in my localhost LM Studio：https://lmstudio.ai API, via uvx (no LLM installation required) and mixing in the llm-tools-quickjs：https://github.com/simonw/llm-tools-quickjs tool plugin for good measure:

LLM’s Python API previously required you to create a conversation and then send messages to it one at a time. This was an abstraction over the true nature of LLMs, where each request carries a complete history of the messages that came before it. That abstraction started to get in the way for some more advanced cases, so the new release introduces a model.prompt(messages=[]) parameter that can be used like this:

LLM previously returned an iterable sequence of strings from each prompt. This worked great when models returned a string response, but failed to predict the weird shape that models would evolve towards. Today many models return a mix of reasoning text, output strings, tool calls, and even image attachments. With LLM 0.32 you can do this instead：https://llm.datasette.io/en/stable/python-api.html#structured-messages-and-streaming-events:

Combine these features and we can finally provide a robust implementation of the semi-standard OpenAI chat completions API, which I’ve now released as the llm-chat-completions-server：https://github.com/simonw/llm-chat-completions-server plugin:

Now you can run prompts against LLM via that server, using the new llm openai endpoint command!

The bigger challenge with that kind of API concerns logging. If we’re going to support the pattern where the message sequence is appended to on every request, ideally we can avoid logging all of that duplicate JSON for every turn.

The solution is the new content-addressable message store：https://llm.datasette.io/en/stable/logging.html#the-message-store, modeled after Git. You can see the new schema for that in the documentation：https://llm.datasette.io/en/stable/logging.html#sql-schema, but the llm logs and llm logs --json commands have both been upgraded to convert that format back into something that’s easy to consume.

There is a whole lot more in this release. The 0.32 release notes：https://llm.datasette.io/en/stable/changelog.html#v0-32 are pretty comprehensive, and the notes for 0.32rc2：https://llm.datasette.io/en/stable/changelog.html#rc2-2026-07-30, 0.32rc：https://llm.datasette.io/en/stable/changelog.html#rc1-2026-07-30, 0.32a3：https://llm.datasette.io/en/stable/changelog.html#a3-2026-06-09, 0.32a2：https://llm.datasette.io/en/stable/changelog.html#a2-2026-05-12, and 0.32a0：https://llm.datasette.io/en/stable/changelog.html#a0-2026-04-28 should fill in any gaps.

Existing LLM plugins should all continue to work, but plugins that provide extra models will need to be upgraded to 0.32 in order to participate fully in the new streaming events system. There’s a guide to implementing plugins with Structured messages and streaming events：https://llm.datasette.io/en/stable/plugins/advanced-model-plugins.html#structured-messages-and-streaming-events in the documentation.

I’ve updated some of my own plugins:

Quite a few of the lower-level tools changes in this release were driven by the needs of Datasette Agent：https://agent.datasette.io/. When I started work on LLM, the term “agent” had such a vague definition that I refused to use it. In September 2025：https://simonwillison.net/2025/Sep/18/agents/ I came around to the idea that " An LLM agent runs tools in a loop to achieve a goal " is well established enough now that I could stop avoiding the term entirely.

Tool chains can now pause for human approval：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-pause and resume from a stored message history：https://llm.datasette.io/en/stable/python-api.html#python-api-tools-resume—both needed by Datasette Agent.

Looking at LLM today it’s beginning to look very agent-shaped to me. There’s something neat about having a CLI utility that can mix and match different tools from different sources with different models all as a one-liner, and that includes a Python library powerful enough to build systems like Datasette Agent：https://agent.datasette.io/ and llm-coding-agent：https://github.com/simonw/llm-coding-agent.

Maybe the next version of LLM will bake the concept of an “agent” into the core library. I’m still trying to figure out what that would look like.

This is New release of LLM adds support for reasoning traces, OpenAI Responses, server-side tools, and smarter logging by Simon Willison, posted on 4th August 2026：/2026/Aug/4/.

Part of series New releases of LLM：/series/llm-releases/

Previous: Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)：/2026/Jul/31/stateless-mcp/

Sponsor me for $10/month and get a curated email digest of the month's most important LLM developments.