gitlearnos
面向 Git 原生 AI 代理的基于证据的学习记忆与自动化
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Guojiz/gitlearnos#1de6f626ab7b6290206663d26de44ce4070b7a5cGitlearnos 以 gitlearnos 发布，当前版本为 2.0.0-alpha.3。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
GitLearnOS comic book Agent wearing an orange Git-branch harness
Learner-owned Git memory · GitLearnOS-exclusive native DeepSeek Harness support
Learning happens anywhere. State belongs to you.
Open the official website → · Core-ready Quickstart →
中文 · Website · Documentation map · Protocol · Author site · X · Bilibili · Sponsor
GitLearnOS gives one capable, replaceable main AI agent a learner-owned Git memory. It notices useful learning events, connects evidence to a goal, guides the next action, and leaves an inspectable, reversible record.
Learning can happen with a teacher, in class, on paper, in a book, on a practice platform, in a project, with peers, or with another AI. GitLearnOS does not move all learning into one application. The main agent connects only the evidence that is useful for the learner's next decision.
The core is ready with a private local Git repository, one write-capable agent, the GitLearnOS protocol, and one subject or real learning event. GitHub and other remotes are optional. During normal learning, Git stays in the background: the learner should not have to manage folders, branches, or commits.
The default recommendation is private. Add a remote only when the learner chooses backup, cross-device continuity, teacher review, collaboration, or publishing, and keep private answers and gaps separate from shared material.
The agent answers the immediate request first. Under safe-auto, it may then
make the smallest safe writeback when the target, evidence, goal, and privacy
boundary are clear. preview proposes the exact change without writing;
manual waits for approval. Original answers, notes, and external feedback
are preserved; AI interpretations remain revisable and must link to evidence.
Once configured, a learner should not need to say “use GitLearnOS” or name a Skill. A question, attempted answer, page photo, class note, teacher comment, practice result, or repeated difficulty can be a learning event. Incidental conversation is not stored.
goal and real input
→ organize traceable evidence
→ ask from the current gap
→ keep the answer or external feedback
→ recheck independently later
→ update state with one reversible Git commit
The success condition is better independent performance on later questions, not a larger pile of generated notes. A normal receipt distinguishes what the current agent did immediately from what a verified scheduled run actually did.
gitlearnos.yml
AGENTS.md
automation.md
dashboard.md
learner-profile.md
subjects/
└── <subject>/
├── goals/
├── sources/
├── knowledge/
├── models/
├── knowledge-gaps/
├── reviews/
└── events/
Only create folders when real learning evidence needs them. Root files hold shared configuration and current views; subject folders hold focused state. Large textbooks, PDFs, scans, media, and long-lived references belong in Project Sources or an authorized local folder. Git keeps compact state, provenance, selected excerpts, and history.
GitLearnOS ships an installable native bundle for the official DeepSeek Harness
Developer Preview. It brings a complete, verifiable Git learning transaction
and an agent-controlled panel into Harness: the main agent owns the ordered
Next up queue and presentation decision, while the learner keeps the final
manual toggle. The Host remains bounded plumbing; it does not invent a ranking
or turn panel state into learning evidence.
The code in this repository currently proves:
learning_status and learning_route bounded, read-only observations;gitlearnos.yml-authorized learning_apply transaction that atomically applies typed
event, knowledge-gap, model, review, and dashboard operations in one
reversible Git commit (with strict learner identity, setup/config, base
revision, and write-authority checks); learning_record remains a
compatibility wrapper;Panel: expand|collapse, and labels development sample data;RAG provider access is a required separate layer for complete deployment; it is
not built into this bundle. A Harness-only installation is therefore
incomplete until a verified RAG receipt is present. DeepSeek's default provider is text-only, but Harness itself is not
limited to text. The learner may either configure a third-party multimodal model
with image input or keep DeepSeek as the main model and install an authorized
vision/OCR bridge plugin. Without either, the agent asks for a transcription
instead of guessing. An immediate multiple-choice answer is supported evidence,
not proof of mastery. Recurring checks use a real scheduler to wake the same main
agent; they do not require a second learning agent. See the
launch note and the
adapter's limits and verification steps.
Complete deployment requires a local or compatible private RAG layer. RAG-Anything is the first explicitly supported option, not a lock-in.
requested
marker is not proof that a run happened. maintenance and due-review are
complete only after each recurring task is created and observed in a real
scheduler.Without verified RAG, the Git learning loop still works as an incomplete
edition. One-off exercises do not enter RAG automatically. If the main agent already understands an image, preserve a
faithful Markdown or structured representation instead of repeating OCR; a
text-only agent must not infer unseen visual content.
Use the core-ready Quickstart, which contains the single canonical setup prompt. It asks the agent to identify the private target, confirm the learning goal, subject, current material, and RAG boundary, wait before learner deployment, detect actual capabilities, and report the undo boundary. The website CTA links to that same source; it does not create a repository or pretend that a button provisioned a scheduler.
The AceSAT demo shows the loop with a fictional learner using short, text-first interactions. It still requires a capable AI runtime; local Git is not the same as a completely offline AI system. Fully offline use would also require a local model and local tooling that the current runtime actually supports.
See the impact statement and the completed SAT fixture for the evidence behind the demo.
unknown.See GITLEARNOS.md for the full behavior contract, QUICKSTART.md for deployment, and Evaluation for documented end-to-end scenarios.
The public face of GitLearnOS is the product site. It carries the same promise as this repository: a live DeepSeek Harness-shaped demo, the canonical Quickstart, and no fake “one-click provisioned scheduler.”
| Website | https://guojiz.github.io/gitlearnos/ |
| Quickstart | QUICKSTART.md |
| Author site | https://guojiz.github.io/ |
| X | https://x.com/guojizh |
| Bilibili | https://space.bilibili.com/3493114115263006 |
| YouTube | https://youtube.com/@guojizh |
| Sponsor | https://github.com/Guojiz/Sponsors |
This branch develops the Git-native v2 protocol and the DeepSeek Harness Developer Preview. MIT License; see LICENSE.