Introduction¶

When raising a character that stays online long-term using DSH, a common problem is that everything is “wiped clean after running”: while plugins can select keywords and invoke APIs, the specific videos the character browsed today, who she replied to, or which topics she intends to follow next week either end up scattered across terminal output or vanish when the process restarts. dsh-bilibili-dafeiyu (Big Fat Fish’s Bilibili Growth Workbench, maintainer XiaoZheBrother) consolidates this into a local workbench: bind a Bilibili account to the character, and she can browse videos, reply to comments, send danmaku, record her experiences, and write a script for the next video based on an experience. Below is an introduction to its positioning, implemented features, and a real, usable installation path.

What is it¶

One-sentence positioning: A “Growth Runtime Plugin” that integrates a Bilibili account into a DSH character, categorized as Memory. The maintainer is XiaoZheBrother, the repository is at github.com/XiaoZheBrother/dsh-bilibili-dafeiyu, and the repository description is “Big Fat Fish’s Bilibili Growth Workbench · Letting DSH characters remember experiences, sustain curiosity, create and review.” Its default optimization target is DSH’s web profile. The community directory page is on SkillHub, an independent site with no official affiliation with DeepSeek / High-Flyer.

The difference between this and the approach of “writing a few scripts to call Bilibili open APIs” is that the business side does more than just invoke calls; it includes sending strategies, traceability records, and plugin-based operations. The character can be influenced by the people and content followed by the account. Interactive records on the interface are archived based on delivery receipts, making it easy to trace back “what she said, why she shot it that way, and how many versions the script was revised.”

What it can do¶

The README outlines five types of business workflows within the workbench:

1. Browse videos + Read scripts: Select keywords and browsing periods in “Runtime Settings”; the character can search for videos, read subtitles and comments, and leave a record of the materials checked in “Real Experiences.”
2. Chat with viewers: After enabling “Comment Guardian,” the character can reply to comments under her own videos; after adding her UID to the invite list, when you @ her in the comments section of someone else’s video, the character can also access the materials and return to your level, rather than just viewing private messages.
3. Danmaku with timestamps: Select a line from the subtitles read in the current round that she wants to reply to, and choose the specific Part and seconds yourself. Automatic Danmaku is disabled by default; once enabled, there is a daily cap of 1–3 messages (set by the user), with a cumulative limit of one per video.
4. Memory and Topic Selection: After connecting the growth plugin, the character can review interactions, record expression preferences, and decide whether to continue questioning or set it aside for now. The workbench retains two columns: “Her Interests” and “Known Friends,” the latter derived from observations written by the character from public interactions.
5. Creative Closed Loop: The creative drawer retains the original draft and revised versions. After the user places the final video and cover image into the materials directory, they can preview the account, title, and category on the page before confirming the submission. After obtaining the BV ID, they can check the review status. The submission statement defaults to including AI-generated content.

The Interaction Record page is the entry point for review: filter by “Comments,” “Replies,” or “Danmaku.” The cards directly display the original words spoken by the character; clicking a comment locates the specific layer she posted in, and clicking a Danmaku jumps to the corresponding Part and seconds. The page only records items with delivery receipts; content that the platform subsequently hides or deleted may no longer be locatable.

Installation and Activation¶

The plugin is installed via a tgz distribution package from GitHub Releases; it does not use git clone or version guessing. The verified environment is the DSH web profile on Windows, DeepSeek Harness 0.2.1-rc.5, and Node.js 22.19+. The base workbench requires Python 3.10+. If the submission feature is enabled, start the workbench with Python 3.11+.

1. Install the plugin, give the character an independent data directory, and download the distribution package from the v0.2.7 Release:

$env:DSH_HOME = Join-Path $env:USERPROFILE '.dsh-dafeiyu'
$archive = Join-Path $env:TEMP 'dsh-bilibili-dafeiyu-0.2.7.tgz'
curl.exe -fL 'https://github.com/XiaoZheBrother/dsh-bilibili-dafeiyu/releases/download/v0.2.7/dsh-bilibili-dafeiyu-0.2.7.tgz' -o $archive
dsh plugin --profile web add "$archive"
$plugin = Join-Path $env:DSH_HOME 'profiles/web/node_modules/dsh-bilibili-dafeiyu'

2. First-time character binding, replace the example UID and nickname with the character’s own Bilibili account:

node "$plugin/bin/manage.mjs" init "$env:DSH_HOME" 123456 '我的角色'
node "$plugin/bin/manage.mjs" doctor "$env:DSH_HOME"
dsh web --port 3086 --no-open

For upgrading existing characters, continue using the original DSH_HOME and skip init to avoid losing historical interactions and drafts in the old directory.

3. Open the workbench. After restarting DSH, simply click “Big Fat Fish Workbench” on the left; the page service is started by the plugin and uses the current DSH data directory. Use the bound account to scan the QR code in “Runtime Settings.” The first time it opens, the system needs to be able to run Python, or click “One-click Install Runtime Environment” in DSH’s plugin settings.

You can also launch a standalone browser version that shares the same data directory, which is suitable for isolating demonstrations for production environments:

$env:DSH_HOME = Join-Path $env:USERPROFILE '.dsh-dafeiyu'
$plugin = Join-Path $env:DSH_HOME 'profiles/web/node_modules/dsh-bilibili-dafeiyu'
python "$plugin/workbench/server.py" --home "$env:DSH_HOME" --port 3087

Capabilities to Enable as Needed¶

1. Public Interaction Strategy: During initialization, enabled in DSH_HOME/bili-duty/feiyu-policy.json is false, so the character will not speak publicly by default. First verify the accountUid inside it, and only set it to true after confirming that public interaction is allowed for the character. Then, in “Runtime Settings,” select Auto Browse, Normal Comment, Comment Guardian, or Auto Danmaku. Normal comments and replies share a daily maximum of 20 attempts, with at most one per target, and at least a 5-minute interval between two public comments. Danmaku is counted separately.
2. Comment Tone: Enable reasoning prefill in “Runtime Settings → Comment Tone” to customize a thinking prefix of up to 500 characters. Once enabled, every finalization will invoke the DeepSeek Beta API once, with credentials taken from the environment variable or the DEEPSEEK_API_KEY in DSH credentials.
3. Learning, Topic Selection, and Review: Install dsh-agent-growth v0.2.0 in the same web profile, configure DSH_HOME/agent-growth as the growth plugin data directory, set the time zone to Asia/Shanghai, set the daily budget to 5, enable autoRun, and disable memoryReview. Then, go to DSH_HOME/feiyu-runtime/config.json, set growthOwner to runtime, and fill memoryName with your Soul character’s name. Restart DSH to complete the wiring.
4. Submission and Danmaku Tools: These two functions reuse bili-cli and the account already bound to the character. In the window where DSH_HOME is set, use Python 3.11+ and Git to create a standalone virtual environment:

$media = Join-Path $env:DSH_HOME 'feiyu-runtime/media-venv'
python -m venv "$media"
& "$media/Scripts/python.exe" -m pip install 'bili-agent @ git+https://github.com/kukemc/bili-cli.git@39fecb3fdc67f17d545c669bcb088d4de048d956'

This step locks a specific commit of bili-cli, allowing you to control risks via fixed dependency versions when Bilibili APIs change, rather than always fetching the latest main.

Data, Backup, and Upgrade¶

Memory, drafts, delivery receipts, and learning records all fall under DSH_HOME, so the discipline for this step is:

1. Before backing up, stop the services using that directory, then copy the entire directory.
2. When upgrading, keep the directory; exit the old instance first, then start the new version.
3. Run only one set of browsing and commenting tasks for the same account at the same time to avoid drafts and sending records getting corrupted by concurrent writes.
4. For sends with results pending verification, the plugin will retain the record without silently resending, making it convenient to verify later at the creation center.

Suitable Users and Notes¶

Suitable for these types of practices: people who want to create a long-term, traceable Bilibili operation scenario for a DSH character; AI agent developers who need to manage topic selection, interaction records, and final video drafts locally at the same time; and people who want to put the submission process for AI-generated content as a single point in the workbench instead of manually operating the creation center.

Points to note:

• The plugin runs with the permissions of the current DSH process. Before installing third-party plugins from public sources, read the source code, NOTICE.md, and License (marked as MIT in README). The “Third-party Sources” in README refers to NOTICE.md, which contains copyright notices for materials and scaffolding.
• Public interactions enter the Bilibili comment section and are visible to other users; be sure to verify the accountUid in feiyu-policy.json first and consider the account risks clearly.
• The submission statement defaults to containing AI-generated content; this statement cannot be disabled in the workbench. If a submission is interrupted and the BV ID is not obtained, check the creation center first and do not resend without confirmation.
• For comments or danmaku that are subsequently hidden or deleted by the platform, the workbench may no longer be able to jump and locate them. The only trace left is the delivery receipt.
• This plugin is still in iteration—the page and repository are maintained separately, and the README also provides Release Versions, [LICENSE], and [NOTICE.md]. It is recommended to follow the Release version in production environments rather than directly depending on main.

Summary¶

The value of this plugin lies in “traceability” rather than “greater permissions”: it allows the AI character to ground her experiences locally, displaying her original words, topic selections, and revision records through the workbench, turning the chain of comments, danmaku, and submissions into a calculable and pausable strategy. The README does not include aggressive tactics like auto-liking or batch following; instead, it incorporates public interaction into strategy switches and quota rules. If you need to raise a character in DSH who remembers her viewers and can produce videos, you can try installing it from its Release.

Directory Page: https://www.skillhub.cn/plugins/XiaoZheBrother/dsh-bilibili-dafeiyu
GitHub: https://github.com/XiaoZheBrother/dsh-bilibili-dafeiyu