StarryDivineSky shows you 256 of its 10,000 links, and the 256 are all one subcategory

The front page is a two month window, not the directory

The repository describes itself as a selection of more than ten thousand projects across machine learning, deep learning, NLP, GNN, recommendation systems, biomedicine, machine vision and front-end and back-end development. The page you land on shows a fraction of that, and the fraction is defined by date rather than by topic.

A tips section states that the README file only shows the first 256 git projects added in the last two months, and that the full content is long enough that it recommends cloning the repository and reading or searching it locally. The rest lives in `CONTENT.md` at the root, next to `README.md`, `.github/` and `LICENSE`, which is the whole repository: four entries, no source tree.

So there are two documents with different jobs. The README is a feed of recent additions, and CONTENT.md is the archive. The default branch is `master` rather than `main`, which matters for anyone scripting a clone. There are no GitHub releases, and the recorded language is unknown, because there is no code here to classify.

Every visible entry sits under one agent subcategory

Read the front page from the entries upward and you find eight projects, and all eight are nested under the same pair of headings: large language dialogue models and data, then agent assistants and robots. They are `deepseek-ai/deepseek-harness`, `bojieli/ai-agent-book`, `TencentCloud/TencentDB-Agent-Memory`, `EKKOLearnAI/hermes-web-ui`, `EvoMap/evolver`, `ValueCell-ai/ClawX`, `GetBindu/Bindu` and `revfactory/harness`.

That is what a two month window does to a list this size. Recent additions cluster, and the window is filled before it reaches hardware or bioinformatics. Meanwhile the table of contents still advertises the full spread, seventeen top level headings running from machine learning and deep learning, NLP, networks and front-end development, machine vision, speech, recommendation, causal inference, finance and time series, reinforcement learning, biomedicine, graph databases, graph neural networks, big data, virtualization and security, through to hardware and a catch-all.

The takeaway is that a reader landing today learns what is new in one corner of the agent tooling scene and nothing at all about the other sixteen categories, which is the strongest argument for cloning and searching instead of scrolling.

An empty heading with no label sits between two categories

The heading structure is not uniform, and one heading is not a heading at all. Directly under the agent assistants and robots section there is a bare five-hash level marker with no text after it, and the entries begin underneath. Whatever category that line was meant to name is absent, and because the table of contents works by anchor, an unnamed level is also an unreachable one.

Naming drifts across the same outline. The first five categories carry numeric prefixes, A01 through A05, glued directly to Chinese text such as `#A01_机器学习与深度学习` and `#A04_机器视觉`. Everything after the fifth drops the prefix. Some anchors are pure Chinese, some are joined pairs like `#强化学习_ReinforcementLearning`, and at least one is a run of characters with no separator, `#图数据库图算法`. Sub-headings follow a second convention, joining the label to a suffix with a separator, as in `其他_机器学习与深度学习` and `神经网络结构搜索_Neural_Architecture_Search`.

Under machine learning and deep learning, eight sub-headings appear with no entries of their own in view: machine learning tutorials, other, distributed machine learning, parameter optimization, anomaly detection, gradient boosting and tree models, feature engineering, and neural architecture search. The same pattern repeats under text generation and dialogue. Whether those sections are empty or simply filled further down the document is the kind of question you answer by searching the clone.

Each entry is a promotional paragraph rather than a summary

Every visible entry carries a long Chinese description, several hundred characters, written in the register of a product announcement rather than a catalogue line. Each one opens by naming a pain point, claims three advantages, and closes with an architectural analogy drawn from kitchens, factories, libraries or city administration.

Those paragraphs also carry numbers nobody in this repository measured. The entry for the genetic programming engine claims event-triggered mutation saves more than ninety percent of energy against periodic full retraining, and that a single agent evolves in milliseconds where reinforcement learning systems need hours. Another entry calls a micro-payment layer for agents a first of its kind. These are the upstream authors' claims, reproduced verbatim in the listing, with no source, no benchmark and no second opinion attached.

This is the practical reason for the 256 item cap and for the clone-and-search advice. At this length, eight entries already fill a screen, so a full document of ten thousand would be unusable as a page. The trade-off is that the directory inherits the tone of whatever it links to, and a reader gets no help telling a measured claim from a marketed one.

A link is the unit of value, and the license field is empty

Because there is no code, there is nothing to install, build or test. The deliverable is a set of links with prose attached, and the value of the repository depends entirely on whether the curation catches things worth having.

The licensing picture needs a direct look. The repository root carries a `LICENSE` file, and the metadata for the project records no license value at all, so the file and the metadata do not agree and neither states a name that can be repeated here. Read the file before you copy anything out of the list; a directory that links to projects does not automatically pass on the terms of the projects it links to, and the list's own terms are the part you can actually check.

The rest of the metadata is thin in the same way. No GitHub releases exist, the primary language is recorded as unknown, and the only stated identity is the project name plus a homepage at wuwenjie.xyz.

The project points outward to a star chart and a Discord server

The visible surface is mostly links leaving the repository. Four GitHub badges at the top go to the issue tracker, the stargazers list, the contributors network and the LICENSE file. A star history chart is embedded from starchart.cc, so growth is tracked through a third party service rather than through anything in the repository.

Community links out as well: a Discord invite sits above the first category heading, and the homepage field points at wuwenjie.xyz rather than at a page inside the repository. The description itself closes with an invitation to star the project and a note that it continues to update, which tells you the maintenance model: one person adding links on a schedule rather than a team reviewing submissions.

The repository is not archived, and its last push is dated 2026-09-06. With no release history and no versioned content, that date is the only signal about whether the list is still being fed, and it is worth rechecking rather than assuming.

Editorial conclusion

Read StarryDivineSky as an index rather than as a survey. It earns its place for people who want a Chinese-language entry point into a field they do not already know, and for the specific links in the newest window. It does not earn a place as a ranking, because the ordering is by addition date, not by merit, and the blurbs are upstream promotional copy rather than evaluation. Before you reuse anything from it, check the license situation yourself: a LICENSE file sits at the root while the repository's license field carries no value, so the file is the only statement of terms. And clone the repository rather than reading the page, since the front page is a 256 item window onto a much larger document.

Frequently asked questions

Official sources

1. Issues
2. Project website
3. README
4. wuwenjie1992/StarryDivineSky on GitHub