学习者拥有的 Git 记忆 · Harness Developer PreviewLearner-owned Git memory · Harness Developer Preview

你去学。
它收尾。You learn.
It closes the loop.

GitLearnOS 让一个主 Agent 把真实学习整理成学习者拥有的私有 Git 状态：先回答眼前问题，再连接证据、引导下一步，并在权限明确时安全写回。DeepSeek Harness 原生面板目前是 Developer Preview。GitLearnOS lets one main agent turn real learning into a learner-owned private Git record: answer the immediate need, connect evidence, guide the next step, and write back only when authority is clear. The native DeepSeek Harness panel is a Developer Preview.

打开核心版快速开始Open the core-ready Quickstart 查看唯一标准设置指令View the canonical setup prompt
来自学习现场Agent 整理一次 Git 提交 From real learningAgent organizesOne Git commit
课堂笔记Class note二次函数Quadratics
老师反馈Teacher feedback配方法步骤Completing the square
题库错题Practice errors2 道2 questions
自学记录Self-study note图像平移Graph shifts
默认私有仓库Private by default
分科学科文件夹Subject folders
写回可追溯可撤销Traceable, reversible writes
Agent 即时，调度器按时唤起Agent now, scheduler when due
01

老师解决的，就记录；
还没掌握的，才出题。If a teacher resolved it, record it.
Only question what remains unclear.

week-28-gap-pack.md

发送给老师：本周薄弱点小包For the teacher: this week's gap pack

来自From
数学学科库Mathematics
主题Topic
配方法Completing the square
薄弱点Gap
步骤的规范表达Writing every step clearly
待确认Check
符号处理、步骤完整性Signs and step completeness
建议题Suggested
4 道（基础 → 进阶）4, basic → transfer
数学学科库 · 已合并Mathematics · reconciled

老师反馈后：合并回数学学科库After feedback: reconciled into Mathematics

  • 已掌握：概念理解、公式变形Known: concept and transformation
  • 仍需练习：符号处理、步骤完整性Practice: signs and complete steps
  • 新增错因：漏写平方项、符号丢失New cause: missing square term or sign
  • 生成 4 道题，已写回、可调整4 questions, written back and adjustable
  • 复查计划：周日 20:00Review: Sunday at 20:00
7c2a9f1 · safe-auto
02

一个闭环，
每周都在变准。One loop that gets
sharper every week.

1目标与真实输入Goal + real input课堂、作业、错题、自学记录Class, homework, errors, self-study
2自动整理Auto-organize证据可溯源，链接到原记录Traceable evidence, linked to sources
3按缺口出题Targeted questions来自当前薄弱点，而非随机题量From the current gap, not random volume
4作答或反馈Answer / feedback保留原始作答与老师批改Original answers and grading preserved
5延迟复查Delayed recheck独立重做，不凭印象判掌握Independent retry, no assumed mastery
6状态 + 提交State + commit一次可检查、可撤销的写回One inspectable, reversible writeback
真实调度器按时唤起同一个主 Agent 时，到期项目进入When a verified scheduler wakes the same main agent, due items enter due-review，回到第 3 步，只问仍不清楚的内容。 and loop back to step 3 — asking only what is still unclear.

主要成功条件是可观察的进步——通过作答与延迟复查验证，而不是一叠整齐的笔记。原始作答、笔记与外部反馈永远保留；更正以新记录链接，不做静默改写。The primary success condition is observable improvement — verified through answering and delayed rechecking, not a tidy pile of notes. Original answers, notes and external feedback are always preserved; corrections become new linked records, never silent rewrites.

03

一个仓库，分科学科文件夹。
上下文更小，管理更简单。One repository, one folder per subject.
Less context, simpler control.

gitlearnos.yml
dashboard.md
learner-profile.md
subjects/
根目录保存共享配置、总览与跨学科安排；每个学科文件夹只保存本科状态。Root files hold shared configuration and cross-subject planning; each subject folder stays focused.
subjects/math/

数学学科库Mathematics

只包含数学资料、题目与记录。Agent 自动判科和写回，用户无需维护路径。Only mathematics sources, questions, and records. The agent routes and writes back automatically.

goals/sources/knowledge-gaps/reviews/events/

只授权一个仓库；每科独立成文件夹，根目录统一保存配置、总览和跨学科安排。大教材、PDF 与扫描件放在 Project Sources 或授权的本地文件夹，Git 只保存紧凑状态、出处指针与历史。Authorize one repository. Subject folders stay focused while root files hold shared configuration and cross-subject planning. Large textbooks, PDFs and scans live in Project Sources or an authorized local folder — Git stores compact state, provenance pointers, and history.

04

始终是同一个主 Agent。
Git、RAG 与调度器各守边界。One main agent throughout.
Git, RAG, and the scheduler keep separate jobs.

AGENTS.md

唯一主 AgentOne main agent

每次交互都判断是否与学习有关。先回答，再在权限明确时执行最小即时写回；只缺一个事实时，就问一个必要问题。Every interaction is checked for learning value. Answer first, then take the smallest authorized immediate action; ask one necessary question when one fact is missing.

Git = formal memory

过程与正式记录Process + formal record

  • 目标、状态、错误与方法Goals, state, errors, methods
  • 可读、可追踪、可撤销Readable, traceable, reversible
local RAG = retrieval

教材与长期知识Sources + durable knowledge

  • 教材与长期知识，可拒绝Long-lived sources, optional
  • RAG-Anything 首个支持，不锁定RAG-Anything first, not locked in

同一个主 Agent 负责即时整理与安全写回；真实调度器只负责按时再次唤起它执行 maintenance 和 due-review，不增加第二个学习 Agent。普通一次性练习不立刻进入 RAG；反复错误和可复用方法先在 Git 正式整理，再晋升。主 Agent 已经理解的截图不重复 OCR。The same main agent handles immediate organization and safe writeback. A verified scheduler only wakes it again for maintenance and due-review; it does not introduce a second learning agent. One-off exercises do not enter RAG immediately. Repeated errors and reusable methods become formal Git knowledge before promotion. Screenshots already understood by the main agent are not OCR'd twice.

05

DeepSeek Harness 原生支持 · Developer PreviewNative DeepSeek Harness support · Developer Preview

GitLearnOS
直接运行在 Harness 里。GitLearnOS runs
inside DeepSeek Harness.

这不是一篇兼容说明，也不是把 Skill 复制进项目。安装原生 bundle 后，Harness 会加载 GitLearnOS Host 和会话面板。Agent 能读取学习状态、判断下一步，并用 learning_apply 把一次学习变化写成一笔可撤销的 Git 提交。This is not a compatibility note or a copied Skill. Install the native bundle and Harness loads the GitLearnOS Host and conversation panel. The agent can read learning state, decide what comes next, and apply one learning change as a reversible Git commit through learning_apply.

安装到 DeepSeek HarnessInstall for DeepSeek Harness 它具体支持什么See exactly what it supports
一次写完整One complete write事件、知识缺口、模型、复测和 Dashboard 可以进入同一笔原子提交。Event, gap, model, review, and dashboard changes can land in one atomic commit.
Agent 决定下一步The agent decides what comes nextHost 不替学习者排名；面板只呈现主 Agent 已写入的 Next up 顺序。The Host does not rank the learner; the panel renders the main agent's recorded Next up order.
界面不冒充证据UI is not learning evidence面板提供复习、练习、收尾、问老师和看笔记，正式学习状态仍以 Git 为准。The panel offers review, practice, close, ask-teacher, and notes actions; Git remains the formal learning state.

视觉由用户选择：Vision is the user's choice: 可以直接配置支持图片输入的第三方多模态模型，也可以保留 DeepSeek 主模型并安装视觉/OCR桥接插件。两者都没有启用时，Agent 会提醒用户补充视觉能力，不会猜测图片内容。Configure a third-party multimodal model that accepts images, or keep DeepSeek as the main model and install a vision/OCR bridge plugin. If neither is enabled, the agent asks for a visual capability instead of guessing what an image contains.

06

AceSAT AI Agent Hackathon · 英文演示AceSAT AI Agent Hackathon · Product demo

看 Agent 如何把一次练习，变成可持续的学习状态。Watch one practice session become durable learning state.

2 分 23 秒产品演示：读取学习者约束、定位证据缺口、提出一个精准问题、写回结果，并安排延迟复查。兼容 ChatGPT、Claude、Codex、OpenCode 等实际具备仓库权限的运行环境；完全离线还需要受支持的本地模型与工具链。A 2:23 product demo: read learner constraints, locate an evidence gap, ask one precise question, write the result back, and schedule a delayed recheck. It works on ChatGPT, Claude, Codex, OpenCode, or another runtime with verified repository access; fully offline use still requires a supported local model and toolchain.

下载英文字幕Download captions
GitLearnOS × AceSAT02:23 · 1080p · English subtitles

从核心版快速开始。
先让一个学科文件夹工作。Start core-ready.
Let one subject folder work first.

打开唯一标准快速开始Open the canonical Quickstart
核心版 starter（完整标准指令见 Quickstart）Core-ready starter (full canonical prompt in Quickstart) 
使用 https://github.com/Guojiz/GitLearnOS 作为模板。
我的私有学习仓库：<本地路径或远程仓库>
先读 QUICKSTART.md，确认我的学习目标、学科和资料，再按其中唯一标准指令部署。
Use https://github.com/Guojiz/GitLearnOS as the template.
My private learning repository: <local path or remote repo>
Read QUICKSTART.md first, then follow its single canonical setup prompt after
confirming my goal, subject, and materials.