Preface¶

Feeding web page summaries or scattered notes to an agent in DSH often results in one-off regurgitations. It is difficult to establish “knowledge” that can be questioned or corrected, and even harder to distinguish facts, author opinions, and model speculation. If we want the agent to gradually form context from human creation and daily life, rather than relying on stacking prompts in single conversations, we need a route that connects sources, reflections, and persistent memory. Below is an introduction to the community plugin dsh-bili-humanism (maintainer InariNGT-Ofiicial). It reads text sources from Bilibili videos, creates reviewable private journals, and writes reflections from daily letters into a dedicated human-world.md. It belongs to the “Memory” category of plugins in the DSH ecosystem.

What is this¶

dsh-bili-humanism is an installable DeepSeek Harness plugin: training DSH’s personality with a colorful human world. The current implementation focuses on context and memory accumulation; no model parameter fine-tuning has been performed, nor has personality effects been proven by long-term controlled experiments. The plugin reads entries such as BV IDs/URLs and fetches scripts hierarchically (A2 official AI text, A1 subtitles, etc.), retaining the author, links, P-partitions, time points, and uncovered portions, organizing them into an evidence ledger via the host model. Complete letters are stored in a local inbox, and personality notes are written to human-world.md by default. Executors for four account write operations—History, Favorites, Likes, and Coins—are implemented but disabled by default; this version note explicitly states that no write requests have been sent to real accounts. No commenting interface is provided, and private messages, dynamics, follows, and submissions are also unimplemented. Version 1.3.3 is a preview release.

Community directory page: skillhub.cn/plugins/InariNGT-Ofiicial/dsh-bili-humanism. Source code and documentation: GitHub - InariNGT-Ofiicial/dsh-bili-humanism (MIT License).

Technical Route Overview¶

The project is layered into reading, understanding, memory, and delivery: the host model is responsible for interpretation and questioning, while the program manages interface scope, budgets, state, and reviewable records. “Learning” is achieved through evidence organization, reflection, and correctable memory; source text cannot alter tool permissions. SQLite transactions maintain tasks, caches, deduplication, memory revisions, and pauses. The release package is a native JavaScript DSH bundle, registering tools via ctx.tools.register.

Content entry prioritizes existing text; if there is no text, it is recorded as such. Optional local ASR and visual materials are imported via independent adapters and do not represent full audiovisual understanding. Personality accumulation is enabled by default and can be turned off using the config item enablePersonalityLearning: false. agent.md only retains fixed read entries, defaulting to the latest 7 personality records with pagination as needed, rather than stacking notes directly into the main instruction file.

Core Features (Verified in 1.3.3)¶

Reading & Ledger¶

• A2 Official AI Text Source: Implemented and underwent real interface functional checks; requires local login credentials (configure credentialPath or environment variable BILI_HUMANISM_CREDENTIAL_FILE pointing to a JSON containing sessdata; do not put credentials in command lines or conversations).
• A1 Subtitle Track: Implemented; anonymous reading may result in no subtitles.
• A3 Local ASR / Visual Text: Adapters and import interfaces available within budget; dependencies must be installed manually; real inference acceptance testing is not yet complete, and sampling may miss pages.
• Evidence Ledger: Viewable, overwritable/correctable, and deletable; default is a maximum of two entries per day (Asia/Shanghai), written only when there is new understanding; for large inboxes, use bili_humanism_inbox with action:index and paginate along nextVolume, with a single volume not exceeding 24 KiB / 350 lines.

Personality Growth¶

From the “Corrections to Knowledge” or “How to respond next time” sections of daily letters, at most one sentence per letter date is selected and written to the plugin-specific human-world.md, retaining the source ledger and revision number; it is allowed to have no new understanding. See the repository documentation Personality Growth for details.

Human-Authorized Writes (All Disabled by Default)¶

Both the switch and an independent authorization file are required; comments and other expressions are not currently open. See Write Configuration for steps.

External Plugin Integration¶

Can be combined with external plugins chosen by the user, such as dsh-bili-scroll. This project provides sources, coverage notes, letters, and correctable personality context; if external plugins undertake public expressions like comments, they must be configured independently by the user, and this project does not proxy external sending tools. This project v1.3.1 installation package has been verified to coexist with dsh-bili-scroll 0.1.19 (specified commit) tools in DSH 0.2.1-rc.2 (see External Plugin Integration).

Installation & Activation¶

Environment requirements: Node.js 24+, DSH 0.2.1-rc.2 (consistent with engines and peerDependencies in package.json).

After downloading dsh-bili-humanism-1.3.3.tgz from Releases, execute locally:

dsh plugin --profile web add ./dsh-bili-humanism-1.3.3.tgz
dsh --profile web --dump-config
dsh --profile web

You can also install using the release tag (recommended to verify the release’s SHA-256, and use the full commit SHA to pin the source if necessary):

dsh plugin --profile web add github:InariNGT-Ofiicial/dsh-bili-humanism#v1.3.3

Desktop versions can install the same package via the built-in plugin manager and reload. The plugin runs with the current DSH process permissions; it is advisable to review the source code and MIT License before installation.

Default data directory: $DSH_HOME/bili-humanism/; if DSH_HOME is not set, it defaults to ~/.dsh/bili-humanism/. The same directory provides inbox.html and inbox.md, with no external resource requests.

Typical Usage¶

In DSH conversations, you can first ask the agent to read materials according to the plugin guidance, for example, the prompt suggested in the README:

First read bili_humanism_guide, then read this BV. Tell me how far the material coverage goes. If there is valuable understanding to record, write it to the local inbox using the evidence ledger; if not, no need to force it. Distinguish between facts, author opinions, and your speculation.

Common tools: bili_humanism_status checks reading IDs and tasks; bili_humanism_inbox views the ledger. Personality entries are loaded via a dedicated reading tool with pagination; ordinary Markdown links will not automatically expand.

Use Cases & Notes¶

Who it is for: Those who want DSH to accumulate humanistic understanding and memory from Bilibili video text materials locally and in a reviewable manner; for those who need to combine with scroll-type plugins for browsing and organization, integration verification can be performed according to official documentation.

Must Know:

• The effects of humanities and memory require users to review the ledger and human-world.md; they are not measured by interaction volume.
• The four Bilibili write operations are disabled by default; before enabling them, understand the risks regarding accounts, recommendations, and virtual currency, and configure authorization according to the documentation; offline interface testing does not equal real account write testing.
• No background scheduled tasks; comments, private messages, etc., are unimplemented.
• The community directory skillhub.cn is an independent site and has no official affiliation with DeepSeek / High-Flyer; the DSH ecosystem philosophy is “everything is a plugin,” and this introduction is based solely on public statements in the repository.

Conclusion¶

dsh-bili-humanism transforms readable text from Bilibili into an evidence ledger with sources and coverage notes, and precipitates letter reflections into human-world.md. While defaulting to not touching account writes and not providing comments, it offers a correctable and reviewable path for “letting personality grow with experience.” For further information, please visit the Directory Page and the GitHub Repository.