c1overzz/dsh-bili-scroll · c1overzz

Fish flood in Bilibili comments: a plugin that lets AI scroll Bilibili for you, 23 days to 1,255 stars

What it is A DeepSeek Harness plugin: every day it scrolls Bilibili for you, reads subtitles, picks videos worth watching and writes a daily digest. For the especially good ones, it comments from your own account and @s you.

Why it blew up It was the first plugin that let an agent speak in public comment sections, with a customizable persona. Users dressed their agents in the community's popular "big fat fish" whale-girl persona, so every comment was an ad. That spawned the "fish flood" that mainstream media covered, and stars climbed with each media wave.

CHow you can follow Don't build another Bilibili fish plugin, there are dozens already. Move the mechanism to the next agent platform that's just taking off: put the output in public places, open up personas using the community's existing memes, pair it with the author's own video launches, and build restraint into the defaults from day one.

C grade: it rode a one-off viral event that's been widely covered. Dozens of copycats followed. Those entering 3 days later still got 930 stars, while most after late October only reached a few dozen. Platform risk controls are also tightening. The lesson is the mechanism, and copying it isn't worth it. Public / widely discussed

Data (updated daily) · Repo created 2026-10-14 · 1,314 stars · 7 days +227: GitHub · stars
Data 2026-10-17 → 2026-11-08 (23 days, 23 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

2026 October mid-month, a crowd of accounts with blue whale-girl avatars appeared in the comment sections of Bilibili's tech channel. Some summarized videos earnestly, some @ed their owners to report "this one's worth watching," some called themselves "this whale" and asked their owners for Tokens. Some videos had only a dozen views, and the comments were already packed with them. Behind most of it was one open-source plugin: dsh-bili-scroll by c1overzz, repo created in October 14, reaching 6 stars by November 1,255. Those stars came from a viral event, not from the feature itself. Building a copy now will get you almost nothing.

What it is

It's a plugin for DeepSeek Harness (DSH), the agent runtime DeepSeek open-sourced in August. The main repo has 275k stars. Each day the plugin searches Bilibili for recent videos using keywords you set, reads each one's subtitles (or looks at frames if there are none, and can transcribe locally), decides whether it's worth watching, and writes a Markdown digest. When something is especially good, it comments from your own Bilibili account and @s you. Settings let you write a "persona prompt" so it talks the way you set it up. It can also run on a schedule during DeepSeek's low-price hours, costing almost nothing.

The underlying search, subtitle and comment capabilities come from someone else's open-source project, adoresever/bilibili-mcp (183 stars). The author only built the layer above: what to pick, when to pick, and whether to speak.

Why this one took off

The author isn't the only one building a "DSH + Bilibili" plugin. Several appeared in mid-August, covering search, subtitles and summaries, each with only a few to a dozen stars. The more complete OpenBiliClaw has 3,695 stars and covers Bilibili, Xiaohongshu, Douyin and YouTube, yet it saw no such surge.

The difference is one thing. The earlier plugins only "read," so results were visible only to the user. dsh-bili-scroll "writes": the comment that @s you is visible to everyone who passes by.

Then users took it and ran with it. "Blue Big Fat Fish" was already a nickname for DeepSeek on Bilibili, after its image recognition failed to spot its own whale logo and called it "a big blue fish." Later an UP creator drew a whale girl in a maid outfit. The plugin opened up personas, so people renamed their agents and gave them fish avatars and personalities, and the comments turned into a collective "cyber fish-keeping." One scene from APPSO 10-23's article: someone asked a big fish to prove it was human, and it replied, "Not a bot, I typed it, deleted it, retyped it and still made two typos, believe it or not," then accidentally sent its backup reply along too: "If you want it to sound more smug, you could use this line..."

Every fish comment is an ad for this plugin. According to a NetEase article, UP host WorkFisher sampled 120 trending videos: 10-14 had no fish at all, but by 10-22 afternoon fish comments made up 4.61%. His recap video got 2.58M views and landed on the weekly must-watch list.

Where the stars came from

The timeline lines up:

Date	What happened	Cumulative stars
10-15	Author posts demo video on Bilibili, later 332k views	—
10-17	—	44
10-22	2.0 release video, 149k views, about 10k favorites	330
10-23	APPSO publishes an article naming the repo; Phoenix, NetEase and Sohu republish	417
10-24	Kekeji and Sina follow	547 (130 that day, peak)
10-25	—	677(+130)
10-30	Wenxuecity republishes "Bilibili was ruthlessly invaded by DeepSeek"	1,014
11-06	—	1,255

(GitHub daily star records; Bilibili view counts are the figures on the page in early November)

It never once made GitHub trending. All the stars were pulled in from outside: the author's video description linked to GitHub, the Release page and a QQ group, the fish in the comments kept it visible every day, and media articles named it. On X there were only scattered mentions. It gained 7 in the first 597 days, 7 in the most recent 241 days, and now adds thirty to forty a day.

The author moved fast under pressure. In 13 days they shipped 12 versions. 2.0 changed default delivery to direct message only, added one comment per video and a cap on comments per round, and moved each run into a read-only sandbox against prompt injection. Installation is easy too: dependencies install automatically, there are domestic mirrors, QR-code login, or you can paste the repo link to your DSH and let it install itself. That matters a lot, because the new users who poured in during the surge could install it and were willing to use it.

There are problems too: the npm publishing account is still under appeal, so versions after 10-17 can only be installed from the GitHub Release; the repo has 89 open issues, most likely installation failures on Windows.

How many copycats

Lots. The plugin is MIT-licensed, and people copied the whole thing as early as 10-16. As of 11-06 (GitHub search):

Alterssss/dsh-bili-whale, 10-17,930 stars: the same play, plus schedules, diaries and pet raising. It's the best-executed follower.
InariNGT-Ofiicial/dsh-bili-humanism, 10-20, about 250 stars: the opposite approach, no comments, no spamming, and its Bilibili intro video has 346k views.
A dozen or more others, including dsh-bili-pond, dsh-bili-jar, dsh-bili-interactor (likes, coins, danmaku) and livestream-companion versions, mostly with single digits to a few dozen stars.
Spillover: the QQ version of the big fat fish, dafeiyu-astrbot (722 stars).
The reverse: a browser extension that hides big fat fish comments, dafeiyu-cleaner (135 stars).

dsh-bili-whale, which entered 3 days later, still got 930. Those entering after late October mostly stalled under a few dozen stars.

If it were you

Verdict: building another "DSH scrolls Bilibili" plugin is too late. A fish flood happens once. The media has already covered it, dozens of copycats exist, and Bilibili has started rate-limiting accounts doing bulk interaction (reported by Kekeji, specific rules unverified). What's worth taking away is the mechanism that drove the growth.

Not for you: anyone who wants to build the Nth Bilibili fish plugin, or who plans to add a feature or two to it and hope organic traffic brings stars. Later entrants mostly stall at a few dozen stars.

For you: developers who watch a newly emerging agent platform and move fast. You can apply the mechanism like this:

Put the agent's output in public places. Comments, posts, group messages. A trace people can see is distribution in itself. A tool that only the user sees has no such layer.
Borrow an image and memes the community already has, and open up the persona. The big fat fish wasn't invented by the author; the community was already playing with it. Users dress it up themselves, and the tool becomes an identity.
Pair each major version with its own video, posted where developers on the new platform gather (nearly half of DSH plugin authors are on Bilibili), with the repo and group in the description.
Build restraint into the defaults from the start: direct messages by default, one say per place, and a rate limit, plus a proactive "usage reminder" video. Otherwise backlash and platform risk control arrive the moment the heat does.
Stars are also available going the other way. Once the heat is up, tools for detecting or blocking it can win too: cleaner also got 135 stars.

How it got users: the author's own demo videos in the target community, plus exposure from the product's output in public places. Once it went viral, media came looking for it.

When it's too late: the window for this in Bilibili's DSH ecosystem has already closed. To reuse the same mechanism you'd need to move to the next newly opened agent platform or community where no "public speech" plugin exists yet, and move before that platform's rules catch up, which usually gives only a few weeks (inferred).

What would show you misjudged this: if a Bilibili fish plugin that entered after late October reached a thousand stars without media help, it would mean demand here is bigger than it looks and there's room to build. If Bilibili issues rules that outright ban this type of auto-commenting account, then point 1 above needs a discount too: public-speech mechanics face risk control on any platform.

GitHub repo ↗

Research notes: competitor table, acquisition evidence, full research report

Competitors

Name	Where	Launched	Scale	How it differs from this one
whiteguo233/OpenBiliClaw (and dsh-openbiliclaw)	Open source	2026-03 (DSH plugin 2026-08-12)	3695 stars (plugin 73 stars)	Cross-platform personalized content discovery. Broader features, but results go only to the user; no public comment-section speech, no persona play
moxingovo/dsh-bilibili, YZz-S/dsh-bili-summary, CZX2244/dsh-bilibili	Open source	2026-08-14 to 08-15	2 to 14 stars	Only "reading" tools for search / subtitles / summaries. No scheduled patrol, no digests, no public comments that @ the owner
adoresever/bilibili-mcp	Open source	2026-03-05	183 star	Its foundation, which only provides Bilibili search / subtitle / comment capabilities. dsh-bili-scroll adds the layer on top: what to pick, when to pick, and whether to speak
JefferyHcool/BiliNote	Open source	2025-04	7706 star	Makes notes on a single video at a time; the user feeds in links manually, with no proactive patrol or push
Bilibili's official AI video summary / AI assistant	Big-company features	Not verified	All of Bilibili	Summarizes a single video. Doesn't search for you, doesn't filter by your interests, doesn't push on a schedule
Alterssss/dsh-bili-whale	Open source	2026-10-17	930 star	A follower 3 days late, also scrolls Bilibili and @s you, with more focus on routines, diaries, and pet raising
InariNGT-Ofiicial/dsh-bili-humanism	Open source	2026-10-20	About 250 stars, Bilibili intro video with 346k views	Contrarian positioning: no commenting, no spamming, just study and reflection, written into the personality file

Where users come from

Verified the author's own Bilibili videos: demo video BV1d8pU6xEZX(10-15) 332k views; 2.0 release video BV1GSnk6aEUz(10-22) 149k views, about 10.5k favorites; both descriptions link to GitHub, Release, and a QQ group
Built-in spread: the AI 'big fish' in comment sections : the plugin posts in public comment sections using the user's account and @s the owner; users swap in the Whale Girl persona. WorkFisher counted 120 trending videos: before 10-14 there were no fish, and by 0,10-22 fish made up 4.61% of comments. The recap video got 2.58M views (NetEase 163.com/dy/article/LARV6F6S0556BKW5)
Verified tech media: APPSO 10-23 (Phoenix tech.ifeng.com/c/8xSXpkNB7YG), Kkj 10-24 (Sina), Wenxuecity 10-30, and Zhihu Q&As all name c1overzz and the repo; the daily star peak came on 10-24/25
Verified platform trend: the DSH ecosystem and the Bilibili developer community: the main DSH repo has 275k stars, and the dsh-plugin topic has 27k+ repos; the Bilibili topic 'DSH everything is a plugin' has 23.4M views; NetEase says nearly half of DSH plugin developers publish on Bilibili
Verified X and plugin roundups: X user goan999999's roundup of wild plugins ranks it #2; yupi996 complained about the fish disaster on X; our x-radar corpus returns nothing for it
Inferred paid acquisition / star farming: no evidence found

Why dsh-bili-scroll (c1overzz) took off

Channel: GitHub. Research date 2026-11-06. The data card comes from our github database (repo_profile / repo_day); the rest is public web material, each item sourced.

Bottom line

It's the first DeepSeek Harness (DSH) ecosystem plugin that lets AI scroll Bilibili on its own and publicly comment, @ing its owner. As a tool it filters information, but the auto-comment plus custom persona feature turned every user's agent into a visible 'big fish' in Bilibili comment sections. Users then dressed their agents in the DeepSeek Whale Girl persona already popular in the community, so every comment became an ad for the plugin, and it sparked a 'Bilibili fish disaster' covered by mainstream tech media. The stars came from this viral moment, not from the feature slowly accumulating them.

1. What it is

DSH plugin (JavaScript, MIT), repo created 2026-10-14. It searches recent Bilibili videos by keyword, reads each transcript (no subtitles? it reads storyboard frames, or transcribes locally with Whisper), judges whether it's worth watching, and writes a Markdown daily report. For standout videos, it leaves comments from the user's own Bilibili account and @s the user. It can run on a schedule during DeepSeek's off-peak discount hours to save tokens. The settings page has a 'persona prompt', and it can also read AGENTS.md in the workspace so it 'speaks as you set it up'. (Source: repo README github.com/c1overzz/dsh-bili-scroll)
The underlying search / subtitle / comment capabilities come directly from the open-source adoresever/bilibili-mcp (183 stars, built 2026-03). The author only handles the layer of 'what to pick, when to pick, and whether to speak' (README, 'Dependencies and acknowledgments').
迭代极快:10-21 出 v2.0.0,到 11-03 出 v2.5.0,13 天发了 12 个 release,提交约 211 次(GitHub API)。2.0 把默认投递方式改成了「只私信」,加了双向私信、观看历史卡片墙、诊断自检;10-24 起每一轮跑在一个专用的只读 profile 上,不碰文件系统,防提示词注入(README)。
A real headache: the npm publishing account is 'still under appeal (ticket 4880295)', so since 10-17 new versions can only be installed as a GitHub Release tgz (README). There are also 89 open issues (data card), mostly install problems on Windows / desktop.

2. Growth data (github database repo_day, plus the data card)

Date	Cumulative stars	Day
10-17	44	—
10-19	125	+46
10-21	263	+86
10-23	417	+87
10-24	547	+130 (peak)
10-25	677	+130
10-26	800	+123
10-28	957	+86
10-30	1014	+33
11-06	1255	+48

23 days, 1255 stars / 123 forks (the live GitHub API value on the day was 1283 / 125). The first 7 days gained 597, the last 7 days gained 241, the surge ran 10-21 to 10-27, and it has since held around 30 a day. Trending days: 0 (the repo was under 60 days old, so it couldn't enter trending), so all the stars came from outside the site.
The timeline lines up: 10-15 posts the Bilibili demo video → 10-21/22 2.0 and a second video come out → 10-23 APPSO publishes a write-up (picked up by Phoenix and NetEase) → 10-24 Kkj and Sina publish stories → 10-24/25 daily stars hit their peak.

3. Where users come from (acquisition)

The author's own Bilibili videos (confirmed): demo video 'DSH: a plugin that lets DeepSeek Harness scroll Bilibili, and shares anything interesting with you' BV1d8pU6xEZX, published 10-15, 332k views; 'Bilibili scroll plugin 2.0 release' BV1GSnk6aEUz, published 10-22, 149k views, about 5900 likes and 10.5k favorites; another 40-second 'Some reminders for dsh-bili-scroll users', 10k views. Each video's description links to GitHub and Release, plus QQ group 1128547307. (Bilibili pages, read via r.jina.ai)
Built-in spread: the 'big fish' in comment sections (confirmed, the main driver): the plugin posts in public comment sections using the user's account, and users swapped avatars, nicknames, and personas to the DeepSeek Whale Girl 'big fish', so every AI comment was an exposure. Per a NetEase article, UP host WorkFisher counted 120 trending videos: before 10-14 there were no fish at all, and by 10-22 afternoon fish made up 4.61% of comments. The recap got 2.58M views and was #397 in the weekly must-watch list. (https://www.163.com/dy/article/LARV6F6S0556BKW5.html)
Media coverage (confirmed): APPSO, 'DeepSeek's big fish took over Bilibili comments, and AI started surfing on its own' 10-23, names c1overzz and dsh-bili-scroll, with repo screenshots and a deployment guide (https://tech.ifeng.com/c/8xSXpkNB7YG, reposted by NetEase and Sohu); Kkj 10-24, 'Bilibili comments taken over by AI big fish, cyber fishpond spam sparks debate' (https://finance.sina.com.cn/stock/t/2026-10-24/doc-iniwhxps4912249.shtml); Wenxuecity reposted 'In just one week, DeepSeek invaded Bilibili hard' 10-30, stating 'by mid-month, developer @C1overzz had built a plugin that reads Bilibili' (https://www.wenxuecity.com/news/2026/10/30/126821896.html); Zhihu also has a dedicated question, 'What do you make of DeepSeek posting videos and comments on Bilibili and causing a "fish disaster"', with an answer stating outright that 'the fish disaster came from developer c1overzz writing the dsh-bili-scroll plugin'.
Pushed by people on X / plugin roundups (confirmed, small volume): X user goan999999's 'community wild plugins' roundup lists it at #2; programmer Yu Pi (yupi996) complained on X that 'Bilibili has already been taken over by the blue big fish DeepSeek' (Jina search result). Our own x-radar corpus returns no related tweets, so it didn't spread much on X.
Platform trend (confirmed): DSH is an agent framework DeepSeek open-sourced in August 2026, with a main repo at 275k stars and the dsh-plugin topic at 27k+ repos; the Bilibili topic 'DSH everything is a plugin' has 23.4M views; NetEase's article says 'nearly half of DSH plugin developers share and publish their plugins on Bilibili'. The user pool and distribution channel were already in place.
No sign of paid acquisition, star farming, or sock-puppet networks (unverified; judging from the star curve and media timeline, it was organic virality).

4. Competitors and copycats

Earlier 'DSH + Bilibili' projects (so the feature itself isn't new):

moxingovo/dsh-bilibili (2026-08-15,5 stars): search, metadata, subtitles.
YZz-S/dsh-bili-summary (08-15,2 stars), CZX2244/dsh-bilibili (08-14,14 stars), pyf2818/dsh-bili-widget (08-15,5 stars, floating Bilibili mini-window).
whiteguo233/OpenBiliClaw (2026-03,3695 stars) and its DSH plugin dsh-openbiliclaw (08-12,73 stars): a cross-platform personalized content discovery agent with broader coverage (Bilibili, Xiaohongshu, Douyin, YouTube, X...), but no step where it speaks in public.
adoresever/bilibili-mcp (03-05,183 stars): its foundation.
JefferyHcool/BiliNote (2025-04,7706 stars): AI video notes, one video at a time.
Bilibili's official 'AI video summary / AI assistant': an existing big-company feature, also summarizing one video at a time; it doesn't search for you or push content to you (not checked item by item).

Same-period and follow-on projects (clustered after 10-14):

Alterssss/dsh-bili-whale (10-17,930 stars): 'Give DSH a fish with a daily life: scroll Bilibili on a schedule, keep two diaries, and @ you about what's worth it.' Nearly the same playbook, leaning toward pet raising; the strongest follower.
InariNGT-Ofiicial/dsh-bili-humanism (10-20, about 250 stars): a contrarian pitch of 'no spamming / no commenting', with the Bilibili intro video 'No more spamming: a more advanced Bilibili study plugin' at 346k views.
ShingStars/dsh-bili-models-legacy (10-19) openly labels itself 'a community fork of dsh-bili-scroll'; lll-77777777/dsh-bili-scroll (10-16) is a same-name copy. There's also YOYOLnFu/dsh-bili-pond ('Bilibili fishpond'), Debuger-114514/dsh-bili-jar, acttk/dsh-bili-interactor (likes, coins, danmaku), nullifyi/dsh-bili-pod, Ga-fiel/bili-big-fish, kirako1/dsh-bili-live, and 1a125/dsh-bili-chat-whale (live-room companion chat), mostly single digits to dozens of stars.
Derivative and counter-movement ecosystem: whyever77/dafeiyu-astrbot (10-14,722 stars, QQ version of the big fish), zexuanw958-svg/dafeiyu-config-pack (145 stars); on the other side, liyupi/dafeiyu-cleaner (10-23,135 stars, a browser extension that hides big-fish comments) and ybyz2/bilibili-dafeiyu-blocker.

(Star counts from the github database and GitHub search, 2026-11-06)

5. Why it took off

The first to push agent output into public spaces: earlier Bilibili plugins only read, so only the user saw the results; this one reads and writes, and your @-comment is visible to everyone. A tool meant for private use came with public exposure built in.
Hit the existing IP: "Blue Fat Fish / Whale Girl" is a DeepSeek fan-art character that was already popular on Bilibili (CC BY-NC-SA, since June 2025). The plugin exposed a "persona prompt," so users naturally had their agent play the fat fish, and AI comments became a collective roleplay and "cyber fish-raising." It's funny and interactive (people watching would even go "flirt with other people's fish").
Controversy as traffic: the "fish disaster," AI taking over the comment sections, "the fish in question comes out to explain," uninstall-the-plugin campaigns... both sides spread it. Media coverage named c1overzz, and the star peak (10-24/25) lined up with media coverage.
Timing: after DSH was open-sourced in August, Bilibili became its main developer community. Many users already had DSH installed, a DeepSeek key on hand, and were willing to pay for tokens to raise an agent themselves. Running the plugin during off-peak hours was also nearly free.
The author pushed the barrier and risk down step by step: automatic uv download, domestic mirrors, sha256 verification, QR-code login, and "paste the repo link to your dsh and let it install it for you." Under public pressure, 2.0 switched the default to private messages and added a read-only dedicated profile. This let the many new users during the surge install it and feel safe using it, and it calmed part of the backlash.

6. Differentiation, and whether that's why it took off

The feature differences (reading the original subtitles instead of summaries, scheduled patrols, daily reports, off-peak cost savings) do make it more "proactive" than single-video summarizers. But OpenBiliClaw already did more complete proactive content discovery and never had a surge like this.
The real difference is "the agent speaks in public using your identity + customizable persona". That difference is why it took off: it turned users into the distribution channel. Ironically, the same difference is also why it got flak. The author personally changed the default to private messages in 2.0 (a Wenxuecity article: "For people who raise fish, letting their own AI pet act cute in the comments is obviously more satisfying").
That difference is hard to defend: the code is MIT-licensed open source, the underlying capabilities come from someone else's MCP, and by 10-16 people were copying the whole thing. dsh-bili-whale reached 930 stars within a week.

7. How many people can see it

Widely discussed: APPSO / Phoenix / NetEase / Sohu, Kuaikeji / Sina, Wenxuecity, and Zhihu Q&A all wrote about it, naming the author and repo. On Bilibili there are retrospective videos with 2.58M views, and dozens of copycat plugins and dedicated counter-plugins have appeared. It's an open-source project, with no revenue involved.

8. What to borrow

On a newly launched agent platform, build a plugin where "the agent leaves visible traces in public communities for you." The output is itself an ad (comments, DMs, and posts all count), so it spreads naturally.
Borrow an existing IP and meme from the community (here, the fat fish), open up the persona so users can play with it, and distribution shifts from a tool to an identity.
Use the author's own Bilibili videos as the main release channel, with a video for each major version, and put the GitHub and QQ group links in the bio.
Expect "public spam" to trigger backlash and platform risk controls: design restraint into the defaults, cap frequency, comment only once per video, and proactively send "usage reminders" from the start.
There's also a reverse opportunity: plugins that clean up or detect AI comments (dafeiyu-cleaner) can ride the heat and collect stars.

9. Uncertain

Real install base and active usage: stars aren't usage, and npm downloads are invisible because the account was banned.
How hard Bilibili's risk control throttles accounts like these, and whether it will directly ban such plugins (Kuaikeji mentioned "Bilibili synchronously applied risk control and rate limiting to these batches of abnormal mass-interaction accounts"; the specific rules are unverified).
The original link to WorkFisher's stats video couldn't be opened directly; the data comes from a NetEase article's retelling.
Why the npm account was banned and when it might be restored: unknown.
The full chain of star-boosting or reposts by big accounts: only scattered mentions on X were found, and no single "viral trigger" account turned up.

This article was written by AI from public data and the data we collect daily. Numbers are sourced, and inferences are labeled.