# 12 hours, 47 directories, 10 days, 326 stars: how many of AgentCompare's stars were earned by the product? · GitHub open source

> Source: https://stringers.io/en/github/gh_lijiaqi-AI-Product-Manager__dsh-agent-compare.html (Stringers / 什么值得做, generated 2026-11-20)
> Stringers tracks which small products suddenly took off each day, investigates why one by one, and distils playbooks from them. Numbers come from 15 databases we collect daily, each with source and date.
> For agents: these investigations and the underlying databases (verified revenue, App Store / Google Play charts, ad libraries, GitHub, Product Hunt…) are queryable live over MCP — tools feed (daily breakouts) / plays (playbooks) / hot / products / data. Setup: https://github.com/komako-workshop/stringers (the user signs in with Google at stringers.io/en/start.html to get a personal URL). Site index: https://stringers.io/llms.txt

> The author used an agent to submit it to 47 DSH plugin directories within 12 hours, including 1024Store built into the desktop app. The remaining stars are likely bought: the forks come from farm accounts with template bios, and BugGPS, promoted the same way, has only 1 stars.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-11

lijiaqi-AI-Product-Manager/dsh-agent-compare · lijiaqi-AI-Product-Manager

# 12 hours, 47 directories, 10 days, 326 stars: how many of AgentCompare's stars were earned by the product?

**What it is** A DeepSeek Harness plugin. For a single request, it has 2–4 candidates each actually run a version, shows the pages, answers, or files side by side in the main chat, reveals the models after a blind pick, and writes back to the repo after review.

**Why it took off** The author used an agent to submit it to 47 DSH plugin directories within 12 hours, including the built-in 1024Store in the desktop client. The remaining stars are likely bought: the forks all come from template-bio farm accounts, and BugGPS, pushed with the same method, has only 1 stars.

**D** **How you can follow it** Don't copy the star-chasing playbook. Take two things: only submit to the directories built into the host client, and track listings with a public issue. The "blind pick, then reveal and explain why" interaction can be moved into your own multi-model tool.

Grade D: the visible success mostly comes from mass directory submissions and suspected star buying. The control, BugGPS, shows that the same submission approach doesn't earn stars. The "same task, multiple solutions, blind pick" gap in DSH is still open, but there's no evidence of any real users, so understanding the playbook is enough. Almost nobody knows about it yet

Data (updated daily) · Repo created 2026-11-02 · 436 stars · 7 days +84

GitHub · stars

Data 2026-11-05 → 2026-11-20 (16 days, 16 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

An AI product intern had a coding agent build a DeepSeek Harness (DSH) plugin for him, and had the agent handle promotion too: 12 hours of directory submissions across 47 plugin directories, and in 10 days it reached 326 stars. But the same person pushed another plugin with the same method and got only 1 stars, and this repo has just 3 forks and 0 followers. Below, I break down its stars: which came from distribution, which were likely bought, and which part is still an open gap worth building for you.

### A PR stopped by a bot

2026 Nov, 4 2:22 UTC: a PR landed on awesome-deepseek-harness, a DSH plugin directory: "Add lijiaqi-AI-Product-Manager/dsh-agent-compare to Tools, Workflows & Presets". 11 seconds later, the directory's anti-spam bot posted a report: the submitter had 0 followers, "14 PRs opened in 100 days", "14 PRs to awesome lists in 7 days", "the same title submitted to 5 repos". By noon that day, a maintainer had closed it for "rapid self-promotion".

A few hours earlier, another directory went the other way. On 11-03 23:52:59, the same account opened a PR to imsai-sh's plugin list; 27 seconds later, a CI bot merged it automatically. That list is the DSH 1024Store, one of the plugin sources built into the DSH desktop app's Discover page.

Put those two together and you have most of the answer to where AgentCompare's stars came from.

### What it is

A DSH plugin (npm package `dsh-agent-compare`, MIT, TypeScript). From the DSH main chat input, click **A|B** , and the same request goes to 2–4 candidates. Each builds in its own git worktree, installs dependencies, and starts its own port. For page-type results, the outputs appear side by side with live preview; for answers and files, they're compared by content. You can vote blind and only see which model and which style produced each version after choosing. Before writing the pick back to the repo, you review the changes, and you can undo after writing. 2.0 adds "say why you picked it": tick a few reasons (layout, color, interaction feel), and later edits in the main chat keep those strengths.

The first line of the README reads: "Same request, 2–4 versions actually run. Pick a style, blind or public compare, then adopt." It's available in 8 languages, with a demo GIF, a row of badges, and a blue liquid-glass UI.

The author's GitHub bio says Xiamen University class of 29, AI product manager intern at Fliggy. The account was registered in 2026 January. The repo itself is almost entirely run by agents: 10 days with 165+ issues and PRs, all opened and merged by the author's own account, with release notes like "All 15 checks passed… the same SHA-256". Versions went from 11-02 (repo creation) through 11-10 to 2.0.0, with 1.0.1, 1.1.0, 1.2.0, and 1.3.0 nearly one per day.

### How well it's growing

| Date (UTC) | stars added that day | Cumulative | npm downloads that day | What happened that day |
|---|---|---|---|---|
| 11-02 | 0 | 1 | – | Repo created |
| 11-03 | +47 | 48 | 478 | first npm package published; directory submissions start that evening |
| 11-04 | +46 | 94 | 380 | 18 awesome-list PRs submitted in a early-morning batch |
| 11-05 | +15 | 109 | 266 | AI auto-writing sites start getting indexed 1.1.0 |
| 11-06 | +48 | 157 | 512 |  |
| 11-07 | +73 | 230 | 1,800 | 1.2.0 |
| 11-08 | +13 | 243 | 5,386 | 1.3.0 |
| 11-09 | +21 | 264 | 7,071 |  |
| 11-10 | +38 | 302 | 9,779 | 2.0.0 |
| 11-11 | +24 | 326 | – |  |

Star counts come from GitHub's daily history; download numbers come from npm download stats.

The download curve looks wild, but the numbers don't add up. The top DSH plugins, dsh-context (about 5K stars) and dsh-free-search (about 1K stars), each bring 4–5 downloads per star per day; on AgentCompare's 2.0 launch day it was about 30 per star per day. At the same time, the install count on the 1024 Store page is 0. Automated traffic is probably mixed into the downloads (inference, not verified).

### Where the stars came from

#### First, it was treated as a one-time batch of directory submissions (confirmed)

From the evening of 11-03 to early 11-04, the author's account opened more than 30 PRs and issues to DSH plugin directories, forked 18 of the directory repos, and filled in web forms on several plugin sites. Issue #4 in the repo contains a table the author maintains. In the 11-05 version it reads: "47 submissions, 14 listed or publicly featured, 1 rejected, 32 pending." Every entry has a status and a link, and it looks like agent-generated and agent-followed-up work.

The heaviest of those 14 listings is the 27-second auto-merged 1024Store from the opening. According to CSDN's description of the DSH desktop plugin marketplace, 1024Store and dshfind are built-in partner sources for the desktop app's Discover page, so desktop users see it when browsing plugins in the app. 11-12, it ranked #3 on the 24-hour star-growth chart in 1024Store's workflow category.

After it was listed, a round of automated content sites picked it up: the Yeyu Piaolin blog (11-04, auto-generating an intro for each plugin), airadar.tv (auto-written articles in 11-05,8 languages, including the line "90 stars"), and dshopen tagged it "🔥 Rising". It looks like people are recommending it, but not one of these was written by a real person. No one on X has mentioned it in nearly 30 days, and Bilibili, Zhihu, Juejin, HN, and Reddit turn up no real reviews.

#### But no amount of directory submissions explains 326 stars

The author 11-04 used the exact same method to submit another plugin, BugGPS (dsh-bug-gps), to more than 15 directories. As of 11-12, BugGPS has 1 stars. His skill-healthcheck and unremember are also at 1 stars each.

Same packaging, same distribution: one got 326, the other 1. The gap of more than 300 stars isn't explained by features.

#### Second, the stars were probably bought (strong inference, can't verify each one)

Several signals point the same way:

- **The forks come from farm accounts.** The repo's 3 forks come from shtickgrisly96, pourquiet4l, and dodgerpax4w. The last two forked within 11-05 21:18 and 21:20, two minutes apart. All three accounts were registered between 2026 8 and October, have 0 followers, and use template bios ("cycling enthusiast. Marketing Specialist by day.", "Somewhere between tennis and reading. Graphic Designer."). Each starred 80–100 small repos, and the repos they starred overlap heavily: a batch of McDonald's MCP contest entries, a codex account switcher, airport reviews. This is the typical client list of a Chinese star-farming network.

- **Another repo by the same author is more obvious.** job_radar has 77 stars and 16 forks, 13 of which come from the same kind of accounts (bios like "Accountant · baking / cycling"), concentrated between 11-05 and 11-08.

- **The ratios are off.** 326 has only 3 forks for 0.9% of its stars, with 0 followers. Comparable DSH plugins created around the same time with similar star counts: dsh-media-desk has 1,145 stars / 129 forks, dsh-game-studio 502 / 44, dsh-live-compact 239 / 18. Older top plugins have fork rates of 4–8%.

- **+47 on day one.** The package was published 11-03, and most directories merge after 11-04 or stay pending, so 47 stars on the same day are hard to explain as organic.

GitHub doesn't let outsiders see which accounts starred this repo one by one, so I can't give a ratio of how many were bought.

#### Third, the DSH plugin ecosystem has a low bar (background)

DSH 2026-08-13 open-sourced. Three months later the main repo is at about 279k stars, the dsh-plugin topic on GitHub already has 30,103 repos, and 4,320 of them were created since November. Awesome lists and plugin sites scrape topics automatically every day, and plenty of their CI pipelines merge PRs straight through. Getting a new repo seen costs almost nothing, which is why 47 spots can be submitted within 12 hours.

### And the idea itself

The feature isn't why it got stars, but the idea really is still an open slot in DSH.

| Competitors | Scale | How it differs from AgentCompare |
|---|---|---|
| Cursor 2.0 multi-agent in parallel | Cursor's main feature | Up to 8 agents run the same task in parallel via worktrees, then you pick; not blind, not in DSH |
| OpenAI Codex cloud best-of-N | Built in | Outputs several versions to choose from; no live side-by-side page preview or reveal |
| LMArena WebDev Arena / Design Arena | Mainstream eval sites | "Two models build the same web page, blind vote, then reveal" is what made them popular; but they only run public evals and don't write back to your repo |
| oil-oil/oil-ui | 3,174 stars, +406 this week | Have an agent output several design directions side by side, then refine; it's a skill, not an independent worktree that runs each version |
| stravu/crystal, claude-squad, vibe-kanban, cc-haha | 3k–29k stars | External workspaces that manage multiple agents and worktrees; no blind pick |
| NanmiCoder/dsh-agent-teams | 2,881 stars | Multi-agent division of labor inside DSH, not multiple solutions to the same task |
| dsh-scalpel, dsh-logos-forge, dsh-recursive-best-of-n, etc. | 0–5 stars | Compare text only, or have a model auto-pick the best |

It's the only one in DSH that brings "run several versions for real → side-by-side page → blind pick and reveal → review, write back, undo" into the main chat. Nobody has copied it yet: 11-07 has a same-named repo, Kyin-code/DSH-agent-compare, which does file diff plus scoring, with 0 stars, and it's a different thing.

Whether anyone wants this slot is the open question, and there's no evidence yet: apart from one outreach bot, the repo has no external issues at all, and 1024Store installs are 0. Cursor and WebDev Arena show that "pick among side-by-side versions" and "blind voting" each have users, but whether anyone wants "blind voting inside DSH on your own work" has never been tested.

### If it were you

**One judgment:** don't copy its star count as a benchmark; most of its stars didn't come from the feature. What you can take are two distribution lessons and one unvalidated idea.

**Who it's for:** developers already building plugins on DSH (or another new agent platform), with a working feature that nobody sees; and people already building multi-model or multi-agent tools who want to add a "blind pick" interaction.

**Who it's not for:** anyone who wants to copy the "12 hours, 47 directories, buy stars" playbook. The denser the submissions, the more likely anti-spam bots will flag you. BugGPS already showed that directory submissions alone don't bring stars. Bought stars come with no forks and no issues, and anyone who knows the space sees through them immediately, so they hurt your résumé. Also not for anyone who wants to build a head-on "AgentCompare 2" inside DSH, since the demand itself hasn't been proven by anyone.

**How to do it:** three things are worth taking.

1. **Only submit to the directories people actually see.** What matters at 47 are the sources built into the host client (DSH's 1024Store and dshfind), plus awesome-dsh-plugin, the de facto catalog with 20K+ stars. When a new platform launches, first find out which sources its client pulls the plugin list from, submit to those, and let the auto-scraping sites pick up the rest.

2. **Track listing status in a public issue.** This practice is good: one row per submission with link, status, and date, so later readers can see where it was listed.

3. **"Blind pick, then reveal" plus "say why you picked it" is an interaction you can reuse.** If you're building multi-model chat, AI design tools, or code generation tools, adding these two steps is worth more than building yet another comparison plugin. If you really want multiple solutions to the same task inside DSH, first test it with 5–10 real users for a week and see whether anyone opens it a second time.

**How to get stars:** don't rely on directory count; rely on a result people will want to share. Record a short video or GIF: "Same request, three models each build a page, blind pick, then reveal", and post it to X and Bilibili. A reveal moment like this spreads naturally (that's how WebDev Arena got its traction). Put the reveal in the first screen of the README, not a wall of badges.

**When to move:** the window of loose DSH directory review and CI auto-merge won't last forever. Some directories have already installed anti-spam bots and reject for "rapid self-promotion"; the further along, the easier batch submissions get blocked. The plugin idea itself has no obvious deadline, but big tools like Cursor and Codex could build blind pick in as a built-in feature at any time.

**Signs that this read is wrong:** ① AgentCompare starts getting real forks and issues from real people, and 1024Store installs climb from 0 — that would mean the idea has real demand and is worth following closely; ② a real person posts a genuine trial on X or Bilibili that brings a wave of organic stars — that would mean "blind pick reveal" itself has spread power and I underestimated it; ③ DSH officially builds multi-version comparison into the main chat — then this slot is gone, and only the interaction is left to borrow.

---

Data and sources: GitHub repo page, daily star and fork records (as of 11-12); npm download stats (as of 11-10); repo issue #4 (directory submission tracker, 11-05 edition); awesome-deepseek-harness PR #211 (11-04, anti-spam bot report and rejection); imsai-sh/awesome-deepseek-harness-plugins PR #807 (11-03, auto-merged); deepseek1024.com public data (11-12); a CSDN article on the DSH desktop plugin marketplace; web search and X over the last 30 days (11-11). "Bought stars" is an inference from the shape and ratio of the fork accounts; I could not verify the starring accounts one by one. npm downloads may include automated traffic.

[GitHub repo ↗](https://github.com/lijiaqi-AI-Product-Manager/dsh-agent-compare)

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

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| Cursor 2.0 multi-agent in parallel | Big-company features | 2025-10 | Cursor's main feature | Multiple agents pick in parallel via worktrees, no blind selection, not in DSH |
| OpenAI Codex cloud best-of-N | Big-company features | 2025 | Built in | Outputs several versions at once, no live side-by-side preview and no blind reveal |
| LMArena WebDev Arena / Design Arena | Web | 2025 | Mainstream eval sites | Blind-votes between two models' web pages and then reveals, but only as public benchmarking and does not write back to your repo |
| oil-oil/oil-ui | Open source | 2026-09-30 | 3,174 stars / 182 forks, +406 this week | Skill where the agent presents several design directions side by side for comparison, does not spin up independent worktrees to run them |
| stravu/crystal(Nimbalyst) | Open source | 2025-06 | 3,128 stars | External desktop orchestrator that runs multiple Claude Code/Codex sessions in parallel worktrees |
| claude-squad / vibe-kanban / worktrunk | Open source | 2025 | 8.8k / 28.7k / 12.5k stars | Manages multiple agents and worktrees, does not do blind comparison |
| NanmiCoder/dsh-agent-teams | Open source | 2026-08-12 | 2,881 stars / 243 forks | Multi-agent division of labor inside DSH, not same-task multi-solution comparison |
| NanmiCoder/cc-haha | Open source | 2026-03 | 15,405 stars | Claude Code desktop workbench with multi-agent, worktree, and diff |
| Limtide/dsh-agent-lab, songrui111/dsh-agent-bench | Open source | 2026-11-02 / 2026-10-26 | 0 / 1 stars | Compares tool-call sequences of two runs for regression testing, aimed at plugin developers |
| elegant01/dsh-scalpel, dsh-recursive-best-of-n, and others | Open source | 2026-08~11 | 0–5 stars | Only compares text answers, or has a model pick the best automatically, with no hands-on trial for the user |

Where users come from

- **Confirmed** **Bulk submissions to DSH plugin directories** : 11-03 17:37–11-04 06:02 UTC, 30+ PRs/issues submitted, 18 directory repos forked; issue #4 self-reports 47 submissions, 14 accepted, 1 rejected for rapid self-promotion

- **Confirmed** **Built-in plugin source in the DSH desktop app (1024Store)** : imsai-sh PR #807 was auto-merged by CI within a minute; the CSDN article says 1024Store and dshfind are built-in partner sources for the DSH Desktop "Discover" page; 11-12 ranks #3 in the workflow category on 24h star growth

- **Confirmed** **Auto-aggregator sites / AI-written content sites** : the Night Rain Drifting blog 11-04, airadar.tv 11-05 (8 languages), dshopen "Rising", dsh-plugin.org, zhengdianxue, and others auto-index it

- Inferred **Bought stars / forks** : 3 forks come from farm accounts with template bios (shtickgrisly96, pourquiet4l, dodgerpax4w); of the 16 forks from the same author's job_radar, 13 are of this kind; fork rate 0.9%, watchers 0; first-day +47; BugGPS, which used the same tactic, has only 1 stars; the stargazer detail endpoint is 404, so I cannot check each one

- Inferred **X / Bilibili / Zhihu / HN / media** : x-radar found 0 posts over 30 days; search turned up no genuine personal recommendations

## Why AgentCompare (lijiaqi-AI-Product-Manager/dsh-agent-compare) could run at all

> Data as of 2026-11-11/12. Sources: our GitHub database (repo_profile / repo_day), GitHub API (repos, PRs, issues, forks, Actions, the author's profile repo data/growth.json), npm API, the public API at deepseek1024.com, Jina / Exa / Brave search, x-radar.

### Bottom line

This is a DSH plugin that an AI product management intern (Xiamen University, GitHub account registered 2026-01) runs fully automatically through a coding agent. Stars came mainly from two things: first, 11-03–04 twelve-hour bursts of directory submissions to 47 DSH plugin directories, including the 1024Store source built into the DSH desktop app; second, stars that were very likely bought. Real demand and usage that can be shown are very small. The product idea itself (running the same request through 2–4 versions, blind selection with reveal, review before adoption) does look like an unclaimed gap in the DSH ecosystem, but it is not where the stars came from.

### What it is

- A plugin for DeepSeek Harness (DSH), the open-source agent framework from DeepSeek 2026-08-13, with 279k stars on the main repo. In the main chat input, click **A|B** , and 2–4 candidates each implement the same request in their own git worktree, installing dependencies and starting their own port. The page version shows live previews side by side (answer version and file version side by side by content). Only after a blind pick are you told which model and which style were used. The chosen version has to pass review before it's written back to the repo, and it can be undone after writing back.

- Release cadence is fast: 11-02 repo created, 11-03 first npm package published, 11-04 1.0.1,11-05 1.1.0,11-07 1.2.0,11-08 1.3.0 (support for plain folders and non-Vite projects), 11-10 2.0.0 (2–4 versions, reasons for picking, blue liquid-glass UI). The README is in 8 languages.

- The whole repo is run by an agent: over 10 days there were more than 165 issues/PRs, all opened and merged by the author's own account, a dozen or more a day, with titles in agent style. For example, the release issue for 1.1.0 reads "All 15 checks passed… SHA-256…", and there is a status table tracking inclusion in 47 directories. The only external issue in the repo (#136) came from an outreach bot from hashgraph-online, tagged `awesome-ai-outreach`.

### Growth data

| Date (UTC) | New stars that day | Cumulative | npm downloads that day | Events in the same period |
|---|---|---|---|---|
| 11-02 | 0 | 1 | – | Repo created |
| 11-03 | +47 | 48 | 478 | 10:41 published the npm 0.0.0-stage; from 17:37 on, it began submitting to directories for inclusion |
| 11-04 | +46 | 94 | 380 | 02:15–06:02 concentrated submission of 18 awesome-list PRs; AI blogs such as the Night Rain Drifting blog auto-wrote posts |
| 11-05 | +15 | 109 | 266 | 1.1.0; airadar.tv auto-indexed it (90 stars at the time) |
| 11-06 | +48 | 157 | 512 | Two farm-account forks |
| 11-07 | +73 | 230 | 1,800 | 1.2.0 |
| 11-08 | +13 | 243 | 5,386 | 1.3.0 |
| 11-09 | +21 | 264 | 7,071 |  |
| 11-10 | +38 | 302 | 9,779 | 2.0.0 |
| 11-11 | +24 | 326 | – |  |

- The daily star counts come from GitHub's `stargazers/history` endpoint, and the author also stored a copy in data/growth.json in the profile repo; the two match. Our database pulls on Beijing time, so dates are one day off: the +96 for 11-08 in our database is the +73 for UTC 11-07 plus part of 11-06. Our database has gain_7d=231, max_day_gain=96(11-08).

- npm downloads come from api.npmjs.org. The top DSH plugin dsh-context has 4,965 stars and about 21k downloads a day; dsh-free-search has 1,077 stars and about 5k downloads a day. That's 4–5 downloads per star per day. AgentCompare was at about 2.0 per star per day when it launched, 30 downloads per star per day. Meanwhile the deepseek1024.com API shows its install count as `installCount: 0`. Much of the npm download traffic is likely machine traffic (unverified).

### Competitors

No other plugin in the DSH ecosystem does the same thing of really running multiple versions, blind selection, and review before adoption, and is gaining traction. The similar tools are agent orchestration tools outside the ecosystem and big-company features.

| Name | Where | Launched | Scale | How it differs from AgentCompare |
|---|---|---|---|---|
| Cursor 2.0 multi-agent in parallel | Big-company features | 2025-10 | Cursor's main feature | Up to 8 agents do the same task in parallel via worktrees and pick afterward, with no blind selection, and it is not a DSH plugin |
| OpenAI Codex cloud best-of-N | Big-company features | 2025 | Built in | Generates several versions at once for you to choose from, with no live side-by-side preview and no blind reveal |
| LMArena WebDev Arena / Design Arena | Web | 2025 | Mainstream eval sites | The "two models build the same web page, blind vote, then reveal" interaction is what made them popular; they are public benchmarks and do not write back to your repo |
| oil-oil/oil-ui | Open source (Claude Code/Codex skill) | 2026-09-30 | 3,174 stars / 182 forks, +406 this week | Has the agent present several design directions side by side for comparison and then refine; it is a design-focused skill and does not spin up independent worktrees to run them |
| stravu/crystal (now Nimbalyst) | Open-source desktop | 2025-06 | 3,128 stars | Runs multiple Claude Code/Codex sessions in parallel worktrees and compares approaches; it is an external orchestrator, not inside DSH's main chat |
| smtg-ai/claude-squad, BloopAI/vibe-kanban, max-sixty/worktrunk | Open source | 2025 | 8.8k / 28.7k / 12.5k stars | Manages multiple agents and worktrees, does not do blind comparison |
| NanmiCoder/dsh-agent-teams | Open source (DSH) | 2026-08-12 | 2,881 stars / 243 forks | Multi-agent division of labor inside DSH, not same-task multi-solution comparison |
| NanmiCoder/cc-haha | Open-source desktop | 2026-03 | 15,405 stars | Claude Code workbench with multi-agent, Git worktree, and diff |
| Limtide/dsh-agent-lab, songrui111/dsh-agent-bench | Open source (DSH) | 2026-11-02 / 10-26 | 0 / 1 stars | Compares tool-call sequences of two runs and does regression testing, aimed at plugin developers |
| elegant01/dsh-scalpel, hua-1111/dsh-logos-forge | Open source (DSH) | 2026-11-05 / 11-11 | 0 stars | One compares two answers within a session, the other rewrites one draft into three versions; both only compare text |
| DSH built-in best-of-N verifier types (dsh-recursive-best-of-n, dsh-turbo, and others) | Open source (DSH) | 2026-08~11 | 0–5 stars | A model picks the best automatically, with no hands-on trial for the user |

Copycats: none so far that copy it. Kyin-code/DSH-agent-compare (11-07,0 stars) has the same name, but it does file/folder diffs plus agent scoring, which is a different thing.

### Where the stars come from: a breakdown of the growth tactics

#### 1. Directory blitz: 12 hours, 47 submissions (confirmed)

- Between 11-03 17:37 and 11-04 06:02 (UTC), the author account opened more than 30 PRs/issues to DSH plugin directories: awesome-dsh-plugin (the de facto directory with over 20,000 stars), imsai-sh (1024Store), bruc3van, 0xsline, vvlife, Herdeny, jiji262, kejixiaoliang, Zhiyuan-Fan, walkinglabs, beancookie, jqueryscript, ZeroPointRepo, dsh-plugin-radar, 2BingLing/dsh-market, polang233, LivXue, and others. It also forked 18 of those directory repos, with fork names carrying the upstream author's suffix, e.g. awesome-dsh-plugin-anil.

- Add the website forms (DSHX, dsh.so, DeepseekPlugin, dshmarketplace.dev, and others), and issue #4 counts itself as 47 submissions, 14 accepted, 1 rejected, 32 pending. The one that rejected it was awesome-deepseekharness #211, with the reason given as "rapid self-promotion".

- The key point: imsai-sh's PR was auto-merged by a CI bot within 1 minutes. This directory is deepseek1024.com, and it's also the Toy "DSH 1024Store plugin chart" on Bilibili (50k plays). According to CSDN's introduction to the DSH Desktop plugin market, DSH 1024Store and dshfind are **built-in partner sources** for the desktop "Discover" page. That means desktop users can see it directly when browsing plugins in the app. 11-12 ranks #3 in the workflow category on 1024Store for star growth over 24 hours, behind archify (96k stars) and dsh-agent-teams (2,881 stars).

- After inclusion, automated content sites amplified it again: the Night Rain Drifting blog (11-04, auto-written intros for each plugin), airadar.tv (auto-written copy in 11-05,8 languages, showing "stars 90"), zhengdianxue, dshopen with a "🔥 Rising" tag, dsh-plugin.org, and others. None of these are genuine personal recommendations. The "third-party blog recommendations" mentioned in recall are, in practice, this kind of automated site.

- **But directory spam alone doesn't explain the stars** : the same author used the exact same playbook to submit BugGPS (dsh-bug-gps) to more than 15 directories, and BugGPS still has only 1 stars as of 11-11. skill-healthcheck and unremember also have 1 stars each. 11-04

#### 2. Probably bought stars (a strong signal, but 100% can't be confirmed)

- 3 of the forks come from farm accounts: shtickgrisly96 (11-03 fork), pourquiet4l, dodgerpax4w (11-05 21:18 and 21:20, forked within two minutes of each other). All were registered 2026-08–10, have 0 followers, and use template bios like "cycling enthusiast. Marketing Specialist by day." or "Somewhere between tennis and reading. Graphic Designer." Each starred 80–100 small repos. The 3 accounts starred 189 repos that overlap heavily: a batch of McDonald's MCP contest entries (mcd-party-planner, mcd-budget-meal, McLife-Desk, etc.), a codex account switcher, airport reviews. This looks like a typical Chinese star-farming network's client list.

- The author's other repo, job_radar (77 stars, 16 forks), makes it even clearer: 13 of its 16 forks are the same accounts (seagull-super5u, uneasytoxic5a, trades62brushed, fusebog3m, argosytruism79, clad-soup0y, clientorgandy8z, stewing-petiole8w, outlets-flops-56……), all with the same template bios like "Accountant · baking / cycling", clustered between 11-05 and 11-08.

- The ratio doesn't add up: 326 stars but only 3 forks (0.9%), with 0 watchers. DSH plugins created around the same time with similar star counts include dsh-media-desk 1,145/129, dsh-marktext 505/50, dsh-game-studio 502/44, and dsh-live-compact 239/18. Fork rates for established top plugins are around 4–8%. Only dbcaskml183-dotcom/dsh-prelude (238 stars, 0 forks, 0 watchers), which appeared the same day, looks similar.

- It got +47:11-03 on day one. The npm package had just been published, and the directories hadn't started reviewing yet (most directories merge or keep submissions pending until 11-04 later). Yet it had 47 stars the same day, so organic exposure is hard to achieve.

- GitHub's stargazer list endpoints (REST and GraphQL) return 404 or nothing for this repo, so we can't check the accounts starring it one by one. This point is a strong inference only.

#### 3. Timing: discovery channels in the DSH plugin ecosystem are still cheap (background, confirmed)

- DSH has been open source for 3 months. The dsh-plugin topic already has 30,103 repos, and 11-01 onward 4,320 new ones were created. Dozens of awesome lists and plugin sites auto-crawl by topic every day, and many directories auto-merge PRs via CI. The bar for a new repo to get seen is almost zero, which is why it can be listed in bulk.

#### Channels not found

- No real person's recommendation or review turned up on X (x-radar 30 days, AgentCompare / dsh-agent-compare 0 posts), Bilibili, Zhihu, Juejin, HN, or Reddit. No media coverage.

### Why (summary)

1. **Self-driven distribution** : the author had an agent submit it to 47 DSH plugin directories and sites within 12 hours. 1024Store is the plugin source built into the DSH desktop app, and this also spawned a batch of automated content sites (the stats table in issue #4, the PR records across directories).

2. **Probably bought stars** : the forks come from farm accounts with template bios. The same author's job_radar has 13 similar forks, a fork rate of 0.9%, 0 watchers, and +47 on launch day. BugGPS, which got the same directory blitz, has only 1 stars.

3. **Platform tailwind** : the DSH plugin ecosystem is in a stage of "many directories, loose review, automatic listing," so a new repo can appear across dozens of lists and the desktop marketplace as soon as it's submitted.

4. **Looks professional** : READMEs in 8 languages, demo GIFs, badges, daily releases, 15 CI checks, and a 2.0 liquid-glass UI make people sent over by the directories more willing to star it. There's no evidence this drove meaningful real usage.

### Differentiation

- Compared with peers in the DSH ecosystem, it's the only plugin that builds "run multiple versions → side-by-side live pages → blind reveal → review before writing back, with undo" into the main conversation. Others compare text only, or only run logs, or let the model pick the winner automatically. Against the parallel multi-agent approach of Cursor and Codex, it adds blind selection and the ability to "say why you picked it, and keep those strengths in later edits."

- This differentiation is **not** why it got its stars. The star curve doesn't match real feedback: apart from the outreach bots, there's not a single external issue, and 1024Store installs are 0. BugGPS, with the same packaging and distribution, didn't take off. What explains its 300 extra stars over BugGPS is suspected star-buying, not features. The idea itself may have real demand (WebDev Arena-style blind evaluation plus Cursor-style parallel runs), but that hasn't been validated yet.

### What others can borrow

- Early in a new platform's ecosystem, treat "directory listing" as a one-time batch job: use an agent to generate each directory's required format (PR, issue form, JSON entry), push dozens at once, and track each listing's status in one public issue. The cost is low. Prioritize sources that are **built into the host client** , like 1024Store and dshfind in the DSH desktop app.

- The two interactions, "blind reveal of the model after choosing" and "explain why you picked it," can be borrowed into any multi-model or multi-agent tool.

- Lesson learned: directory blitzing buys exposure but not stars. BugGPS is the control group.

### Still unclear

- How many of the stars were bought: stargazer details aren't available, so no ratio can be given.

- How much of npm's daily 5–10k downloads are real installs. The cause of the surge starting 11-07 is unclear: it could be automated installs or inflated counts, or a desktop source may have pushed it to the front.

- Whether there was an external event behind the +73 stars on 11-07: none found.

- Whether dbcaskml183-dotcom/dsh-prelude is another client of the same star network: same pattern, not investigated in depth.

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