Pi 1.0

ID: f796b39f
原文链接: https://earendil.com/posts/pi-1-0
作者: Earendil
日期: 2026-10-01
分类: agents
来源类型: article
标签: agent-harness、codemode、mcp、coding-agent
质量评分: 4/5
抓取时间: 2026-10-03 (content-fetcher, opencli web read)

中文导读

Earendil 发布 Pi 1.0：定位为 hardened、极简、可扩展的 agent harness。此前刚以 "You said no MCP" 拒绝默认接入 MCP，这次在保持极简路线前提下加入：Codemode（原生支持 MCP 与 Jev、图像模型等非 LLM 模型）、虚拟模型扩展（官方 demo：Claude Opus 做规划、GPT 6 Luna 做实现、Jev 判断切换时机，注册为 router/auto 模型）、延迟工具加载、Anthropic 模型缓存预热、会话中途系统消息（transcript 感知的 prompt 与工具变更，支持的模型只发送增量以保持 prompt cache 有效）、新 TUI 主题、默认全屏模式。发布哲学：agentic 工具每周都在变但大多不长久，Pi 等某件事被证明有效后才考虑采纳，并权衡真实功能与引入的复杂度。作者称经过数月基于社区 issue/PR 的加固与维护，Pi 已成为个人与企业可依赖的稳定软件，每周有数十万人使用。同日发布实验性新包 Pi Durable 面向长时运行 agent（见相邻条目）。MIT 许可，文档在 pi.dev，代码在 GitHub earendil-works/pi。

为什么值得关注

极简 harness 路线在 1.0 节点同时接纳 Codemode/MCP 与虚拟模型路由，说明"延迟采纳"策略与快速演进的 agent 生态可以兼容，是 harness 设计的活样本。

English Summary

Earendil shipped Pi 1.0, a hardened, minimal, extensible agent harness. Additions on top of its minimal coding-agent core: Codemode (native support for MCP and non-LLM models like Jev and image models), extension support for virtual models (the official demo wires Claude Opus for planning and GPT 6 Luna for implementation with Jev deciding when to switch, registered as a router/auto model), deferred tool loading, cache warming for Anthropic models, mid-conversation system messages (transcript-aware prompt and tool changes, sent as deltas only on models that support mid-conversation changes so prompt caches stay valid), a new TUI theme, and full-screen mode by default. The release philosophy: agentic tooling changes weekly but most changes do not last, so Pi adopts only proven ideas and weighs true functionality against inherent complexity. After months of hardening from community issues and pull requests, Pi serves hundreds of thousands of weekly users. An experimental sibling package, Pi Durable, targets long-running agents (see the adjacent entry). MIT licensed, docs at pi.dev, code at github.com/earendil-works/pi.

Obsidian 原文摘录（抓取正文头部）

# Pi 1.0
> 发布时间: 2026-10-01
> 原文链接: https://earendil.com/posts/pi-1-0

---

# Pi 1.0

日期：2026年10月01日周四

发件人：Earendil <[rfc@earendil.com](mailto:rfc@earendil.com)\>

收件人：您

主题：Pi 1.0

Today we are proudly shipping Pi 1.0: a hardened, minimal, extensible agent harness that you can make your own. Hundreds of thousands of people around the world use Pi every week. Many of you submit issues and pull requests. Over the course of many months, we have used that feedback to improve, harden and evolve Pi into a stable piece of software that people and businesses can depend on.

Pi is known for being minimal. We care about holding that line. While agentic tooling changes every week, many of the changes do not last. Pi does not work that way. We wait until something has proven itself, and only then do we consider adopting it; weighing its true functionality against its inherent added complexity. We spoke more about that process and how it relates to Codemode and MCP earlier this week [here](https://earendil.com/posts/you-said-no-mcp/).

The Pi 1.0 we are shipping today is a result of that process. Pi already ran the latest models from every major provider, became a daily driver coding agent for people around the world, and provided a supermalleable substrate on which to build agentic applications. With Pi 1.0 we are adding the following into Pi:

-   Codemode (native support for MCP, and non-LLM models like Jev and image models)
-   Extension support for virtual models
-   Deferred tool loading
-   Cache warming for anthropic models
-   Mid-conversation system messages (transcript-aware prompt and tool changes)
-   A new TUI theme
-   Full-screen mode by default

We have thought about many of these features for months. They were thrown up against the wall and stuck. The list of things that fell off the wall is much longer. When we use Pi with all these new features it feels like a major step forward, but it also still feels simple, like Pi.

Alongside this work of continually refining and evolving Pi, we also began to understand that there were aspects of Pi that did not fit the shape of how many people wanted to use it. At Earendil, we want to bring Pi's minimalism and the way it allows you to wield AI outside of the coding agent and outside of the terminal. It needs to be reachable from different surfaces and support longer-running conversations and tasks. In short, Pi must become more durable. Rather than stray from our minimal roots and try to make Pi something it wasn't, we combined that work into a new experimental package that we are shipping today, Pi Durable. Pi Durable is a new substrate for building long-running agentic applications that lets the builders and the users of those applications wield and steer the underlying intelligence with tremendous dexterity. It shares Pi's key principles, minimalism and supermalleability, but it extends those principles into new dimensions. Mario writes in more detail about how we crafted Pi Durable [here](https://earendil.com/posts/pi-durable/).

Earendil was formed to craft software and open protocols that strengthen human agency. Today we are shipping tools to you all to help do just that. We're eager to hear feedback from you and to harness the future out in the open together.

Pi 1.0 is available today:

curl -fsSL https://pi.dev/install.sh | sh


or on Windows

powershell -c "irm https://pi.dev/install.ps1 | iex"


Pi Durable is available today as an experimental package:

npm install @earendil-works/pi-durable @earendil-works/pi-ai @earendil-works/chord


Both are MIT licensed. Documentation is at [pi.dev](https://pi.dev/). Code is at [github.com/earendil-works/pi](https://github.com/earendil-works/pi).

                                                                                      v1.0.0                                                                          escape interrupt · ctrl+c/ctrl+d clear/exit · / commands · ! bash ·        ctrl+o more                                                                     Press ctrl+o to show full startup help and loaded resources.                    Pi can explain its own features and look up its docs. Ask it how to use or      extend Pi.                                                                     \[Context\]                                                                         AGENTS.md                                                                     ────────────────────────────────────────────────────────────────────────────────Use co                                                                          ────────────────────────────────────────────────────────────────────────────────~/Development/pi-mono (main)                                                    0.0%/1.1M (auto)                                             gpt-6-luna • medium

00:00\-01:5900:25 - Pi writes itself an extension: a virtual model that plans with Claude Opus and implements with GPT, with Jev deciding when to switch01:01 - Reload, start a new session, and pick the new router/auto model01:14 - Planning runs on Claude Opus01:27 - Jev spots the switch to implementation and hands over to GPT 6 Luna01:44 - ! runs the finished script01:50 - /session breaks down cost per model and cache useKeyboard shortcuts (?)Fullscreen (f)

Notes

Content grounded in an opencli web read fetch of the full article (2026-10-03 (content-fetcher, opencli web read)).
中文导读 is the entry's summary_zh (light cleanup from the fetched source); 为什么值得关注 is the entry's one_liner; no claims beyond the fetched source text were added.
Source file cached at /tmp/aaif_pi1.md during the run.