返回
RCreddit.com
17
·开发者社区 · RSS

open source desktop alternative to codex app and claude code with better features and control

查看原文
推荐理由

这条记录涉及编程工具或代码能力更新,适合开发者评估工作流变化和可复用价值。

AI 摘要

Limboo 是一款开源桌面应用,旨在解决现有 AI 编码工具在大型项目开发中工作流程碎片化的问题。它将 AI 视为工程工作区的一个组件,而非整个应用。Limboo 为每个任务创建独立的会话,包含对话历史、终端状态、Git 分支、检查点和本地内存等。其核心功能“恢复管道”能确保 AI 在会话恢复时,基于验证过的代码库状态而非对话记录继续工作。…

Over the past few months I've been building an open-source desktop application called Limboo, and I wanted to share the idea behind it because I'm curious whether anyone else has been running into the same problems.

One thing I've noticed with AI coding tools is that they're incredibly good at writing code, but once a project becomes large, the actual engineering workflow still feels fragmented.

The AI is usually in one window.

Git is somewhere else.

The terminal is somewhere else.

Build logs are somewhere else.

Documentation is in another browser tab.

Project decisions are spread across old conversations.

The longer I work on something, the more time I spend rebuilding context instead of actually building software.

That observation is what started Limboo.

The goal isn't to replace coding agents like Claude Code, Codex, or other agent-based tools.

I actually want to use those.

The idea is to build everything around them.

Instead of treating the AI as the entire application, Limboo treats it as one component inside a much larger engineering workspace.

Every task becomes its own isolated session.

Each session has its own conversation history, terminal state, Git branch or worktree, checkpoints, permissions, local memory, search index, execution timeline, and task list.

The agent resumes against verified repository reality, not a transcript. The Resume Pipeline is the flagship of this idea. When you reopen a session — after an hour or after three weeks of other people's commits, rebases, and dependency bumps — Limboo revalidates the git worktree against the exact state the session last saw and computes a structured repository delta : commits landed, files changed (with dependency manifests and migrations flagged), symbols added or removed, and which files import what changed. That delta is injected once, before your next prompt, so the agent reconciles its assumptions up front instead of burning turns re-reading the tree. It is fully local, uses only bounded argv-only git, and never blocks you from switching sessions.

If I stop working on a feature today and come back in two weeks, I don't want to explain everything again.

I want to reopen the session and continue exactly where I left off.

Another thing I wanted to improve is transparency.

I don't like when an agent runs commands that disappear into a log somewhere.

If it's running a build, I want to watch the build.

If it's modifying files, I want to see the diffs while it's working.

If it wants approval, I don't want a giant modal that interrupts everything—I want approvals to appear naturally inside the conversation stream.

主题标签Claude开源代码端侧推理
原始关键词#alternative#features#control#desktop#better
查看原文reddit.com
单一来源,暂无交叉验证
open source desktop alternative to codex app and claude code with better features and control · BuzzRadr