# "Why do I have to wait until 80% to compress?" A Bilibili UP-owner's compression plugin, 11 days, 260 stars · GitHub open-source project

> Source: https://stringers.io/en/github/gh_Tranquillitatis-Lab__dsh-live-compact.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

> Right when official compaction was getting heavy criticism, the author launched with one line, "don't wait until 80%, tens of milliseconds, 99% cache hits," plus two self-shot Bilibili videos (about 108K views combined), then kept heavy users with side-by-side benchmarks and daily issue replies.

[GitHub open-source project](https://stringers.io/en/github/index.md) · written 11-13

Tranquillitatis-Lab/dsh-live-compact · Tranquillitatis-Lab

# "Why do I have to wait until 80% to compress?" A Bilibili UP-owner's compression plugin, 11 days, 260 stars

**What it is** A DeepSeek Harness plugin that replaces the official compression. At the end of each turn, it swaps the long output from that turn for an excerpt card on your local machine. It calls no model, leaves the cached prefix untouched, and the original text can still be retrieved.

**Why it took off** It landed in the few days when the official compression was getting widely criticized. The author pushed it out with a one-liner, "no need to wait for 80%, tens of milliseconds, 99% cache," plus two self-shot Bilibili videos (about 108k views combined). Then comparison reviews and daily issue replies kept the heavy users around.

**C** **How you could follow this** Don't build yet another DSH compression plugin. Learn the playbook instead: during the one or two weeks when a platform's built-in feature is being criticized en masse, build a replacement plugin. Title it with a rhetorical question, publish a comparison review on day one, shoot your own video, and a week later post a follow-up that addresses the comments.

C grade: the playbook is clear and worth learning, but the stars came from an official-failure window that is closing and from the author's video channel, which followers may not have. There are already dozens of DSH compression plugins. Even if every condition is met, the ceiling is about two or three hundred stars. Being passed around in small circles

Data (updated daily) · Repo created 2026-11-03 · 309 stars · 7 days +62

GitHub · stars

Data 2026-11-07 → 2026-11-20 (14 days, 14 observations). Points before 10-28 come from deeply / archive backfill, not every day has one

DeepSeek Harness's built-in context compaction had the worst word of mouth in early November: it waited until 80% full, took minutes to compress, wiped the cache on every run, and sometimes looped forever. A Bilibili creator who runs local LLMs posted a plugin right then, with a 16-minute video, and within 11 days it had over 260 stars. The same idea had been built before and stalled at 17 stars. The difference was the video, the benchmarks, and how fast he replied to issues. Below: what it is, where the stars came from, who else is building in this space, and what you'd need to copy the result.

### Night of Nov 3, one video

2026 Nov 3, 21:55 Beijing time: Bilibili creator Merkyor posted a video titled "Deepseek Harness real-time self-compaction plugin Live-Compact open-source intro: cherish your context, stay away from thunderous mega-compaction!" "Thunderous mega-compaction" is his nickname for the official compaction: once the conversation reaches around 80% of the window, DSH stops and calls the model to write a long summary, which can take several minutes, the summary loses details, and the prefix cache built up earlier is wiped out entirely. In the intro he boils his approach down to one line: you can compact small amounts at any time, you don't have to wait until 80%, compaction takes tens of milliseconds, and you can keep a cache hit rate above 99%.

As of 11-14, the video had 59k views, 2,166 likes, 4,197 favorites, and 2,105 comments (Bilibili). Comments were 3.5% of views, which is high. Some comment accounts were clearly AI accounts driven by the DSH plugin; I didn't verify how many.

The timing wasn't random. Around 11-01, the DSH official forum had a run of compaction complaints: the official compaction plugin dropped a parameter, so the prefix cache missed entirely (#11816); once history exceeded the summary limit, `/compact` failed 101 times in a row and looped (#11773); earlier, someone reported "a 120K window compacted at 38K" (#11263). With DeepSeek API prices, the gap between cache hits and misses is large, so on long tasks a single compaction that wipes the cache costs real money.

The repo got its first commit that evening (18:36), the README was revised 10 times that night, comparison charts and slides against the original were added, and then the video went out.

Merkyor (MerkyorLynn on GitHub) describes himself as "an individual developer who likes post-training" and has 5,686 Bilibili followers. His earlier videos covered post-training local Qwen models; "Qwen3.8 27B Cyber Therapy finale" got 112k views. So his first viewers were already people running models locally and sensitive to context length.

### What it actually changes

It's a third-party DSH plugin, MIT-licensed, and installing it replaces the official `dsh-compaction-basic`. The official approach is "wait until nearly full, then compact a big chunk at once." This one changes when and where compaction happens:

- **Cut a little after every turn.** Once context passes 20k tokens, long logs and long tool outputs the model has already read are swapped in place for a "card" that keeps the opening and closing excerpts. This runs locally, calls no model, and takes tens of milliseconds.

- **Never cut backwards.** The cut point stops at the previous card, so prefix content already cached is never touched, and cache hits hold. This rule came from data on a single user's 32-hour long session in issue #19.

- **Originals stay recoverable.** Swapped-out text is backed up locally in `~/.dsh/compaction-spine/shadow` and can be read back by ID.

- **The official fallback still exists.** If the window really fills up, the official large summary still runs once.

- **Cloud and local are handled separately.** People using cloud APIs like DeepSeek or GLM care about cache and cost; people running models locally care about context fill and prefill time. The plugin tells them apart from the configured endpoint. The second group is the creator's original audience.

The README leads with benchmarks: on one real session, both versions answered all 15 recall questions correctly; there's a comparison table of 150 early short tasks and 20 long tasks; and someone else ported it to OpenCode and ran their own comparison, with all Aider 150 and Web-Bench React 20 problems passing, and estimated costs 14.3% and 42.1% lower. Each result says exactly what it measured: historical data is labeled as such, and results from other projects aren't presented as DSH results. For a technical audience, that's more credible than one big number.

### Where the stars came from

GitHub daily stars (day-by-day data available from 11-07):

| Date | star | Added that day |
|---|---|---|
| 11-07 | 113 | — |
| 11-08 | 135 | +22 |
| 11-09 | 142 | +7 |
| 11-10 | 159 | +17 |
| 11-11 | 221 | +62 |
| 11-12 | 236 | +15 |
| 11-13 | 247 | +11 |
| 11-14 | 260 | +13 |

The first four days' 113 stars came mostly from the 11-03 video. Day-by-day gains from 11-03 to 11-06 aren't available, so the exact number the first video drove can't be separated out.

The later peak matches the creator's second video. On the evening of 11-09 22:37 he posted "Keeping model attention as high as possible at the lowest possible context fill — design and implementation of DSH Live Compact": 49k views, 2,762 favorites, 977 comments (Bilibili). The description says "after the last video I got hundreds more comments," and this one was made to answer the questions from the first. The next two days were +17 and +62, the highest of the whole period; npm downloads also peaked at 11-10 in one day, with 969 downloads.

A few smaller channels helped too:

- **A small ecosystem group.** Around 11-01, several DSH plugin authors created a GitHub org, Tranquillitatis-Lab, and the repo moved from the author's personal account into it around 11-07. T-Auto in the org built dsh-ops, which has 5,298 stars; he's also a Bilibili creator and a member of the dsh-TUI team. He filed a compatibility issue on day one (#2) and has made 27 commits to the repo since. The dsh-TUI author left a comment on the video with 48 likes.

- **Others ported it.** A Pi port appeared the day of release (issue #1); a week later someone made an OpenCode version with its own benchmarks (#40). The author put those benchmarks at the top of his README.

- **Directory sites picked it up automatically.** dsh.so ranks it #8 on its benchmark list, awesome-dsh-plugin lists it in the TOP200, and a dozen or so plugin directories and blogs aggregated it automatically. The author didn't pitch any of them.

- **X and the tech press haven't picked it up.** A DSH plugin roundup video hit 11-12 and got a second episode, unverified whether it's monetizing; since then it's only adding about ten a day.

There's also a conversion rate to note: the two videos together drew roughly 108k views and nearly 7,000 favorites, and turned into 260 stars. npm had 11-04 downloads from 11-12 to 5,501 (this includes auto-updates and mirrors, so it isn't a headcount), and the author also runs a domestic mirror that isn't counted. Bilibili viewers mostly install directly and don't star on GitHub. If stars are the goal, Bilibili is an inefficient way to get them.

### Same idea, why someone else got 17 stars

The DSH compaction plugin space is crowded: dozens of repos have compaction or compact in the name, most with one to five stars:

| Project | Launched | Scale (around 11-14) | How it differs |
|---|---|---|---|
| Official dsh-compaction-basic | 08-13 | All DSH users | Compacts a big chunk only when nearly full; live-compact still falls back to it |
| ranxianglei/billion-context | 08-06 | 7,266 stars, 9 M npm downloads in 1.05 days | Top of the category; the model prunes itself; works across Claude Code, Codex, OpenCode, and Pi |
| Tyan66666/billion-context-dsh | 08-13 | 169 star | Same track: model-driven proactive pruning |
| mengwEX/dsh-block-ptc | 10-15 | 103 star | Lets the agent decide when to compact |
| TsFreddie/dsh-compaction-instant | 08-14 | 17 stars, 9 npm downloads in 159 days | Closest in approach: no model call, millisecond-level, recoverable; but still compacts a big chunk at a threshold, doesn't discuss cache, no video or benchmarks |
| mrbeandev/dsh-hypercompact | 09-26 | 3 star | No model call; compacts against a byte budget |
| dsh-compaction-pro / -cacheaware / dsh-smart-compaction / dsh-compaction-tower and others | 08–November | Mostly 1–5 stars | Each changes one thing: summary quality, the threshold, or cache repair; none got any push |

The clearest comparison is dsh-compaction-instant: about two and a half months earlier, with the same no-model approach and the same recoverable originals, it reached 17 stars. "Local extraction, no model call" isn't a new idea. What live-compact added was the combination of small cuts after every turn, never cutting backwards, and preserving the cache, plus one line anyone can understand, a benchmark set people can check, and the author's two videos. The product differences made technical viewers willing to try it and to write long issues, but most of the traction came from the author's distribution and engagement (my inference).

### Where it falls short

- **Stability.** 11 days with 54 issues and PRs and 13 npm releases. Between 11-07 and 11-08, 6 Windows-severity issues came in: permission setup on the backup directory failed and crashed the DSH host process (#20, #22–#25, #28), and the author shipped three releases in one day to fix it. By 11-13 people were still reporting 1.1.6 hard freezes, no cuts at all, and the whole package being silently skipped by DSH (#49, #53, #54). The repo still has 37 open issues (GitHub, 11-14).

- **Heavy users are doing code review for him.** Many issues are reproduction reports thousands of words long, with measured numbers and proposed fixes. Of the 196 commits, 87 are credited to Kimi and 27 to Codex, so the author is using AI agents to keep up. That's part of why people stick around, and it also means code changes fast and regressions are frequent.

- **It only has weight inside DSH.** The Pi and OpenCode ports have 4 and 7 stars. Outside the ecosystem, without the author's videos, the same code gets no attention.

- **The ceiling is low.** Compared with billion-context, the category leader, npm downloads are about two hundred times lower. It's known in the DSH Chinese community and beyond that circle, few people have heard of it.

### If it were you

**Verdict** : C. The project shows a complete playbook, but copying it to build another compaction plugin isn't worth it. Its stars came from two things a follower can rarely replicate: the few days when official compaction was failing in early November (the window was open, and the official side would eventually fix it), and a creator with several thousand followers who could make a 59k-view video. DSH already has dozens of compaction plugins; even with all of that, it only turned into 260 stars.

**Who it suits** : plugin authors who already have a few thousand technical followers on Bilibili or X, can explain the principle themselves, and are willing to answer issues every day; or people hitting the same "built-in feature gets roasted" problem on platforms like Claude Code, Codex, or Kimi Code.

**Who it doesn't suit** : people without a video or posting channel who expect organic GitHub traffic; instant's 17 stars answer that. People who want to build yet another compaction plugin for DSH; that space is already full. People who don't want to deal with edge cases like Windows permissions every day.

**How to do it** , ranked by how worthwhile:

1. **Copy the playbook, and wait for the next "official feature gets roasted" moment** : watch the official DSH, Claude Code, and Codex forums and issue trackers. When a built-in feature gets a week of concentrated complaints, build a replacement plugin, title it with a rhetorical question (this project used "Why wait until 80% to compact?"), publish a side-by-side benchmark against the original on day one, and state the methodology plainly. That's the most valuable part.

2. **Port it to another ecosystem, but bring your own distribution** : the "small cuts each turn, preserve the prefix cache" idea doesn't yet have a well-known version on Claude Code or Codex (my inference; I didn't check each one). A port alone isn't enough; the Pi and OpenCode ports each have single-digit stars. You need a video or long post explaining the principle, with your own cost comparison screenshots.

3. **Build what it doesn't do well** : for example, add a "search original" tool (someone asked for this in #16), manual cut points (#48), or a reusable compaction benchmark suite that anyone can use to compare plugins. You'd be building reputation, and the stars would be fewer.

**How to get users** : post the video first, then push GitHub. The first video covers what it is and why; a week later the second answers the comments from the first, and this project's peak day came after the second. Reply to every comment and issue. Welcome ports from others, and put third-party benchmarks at the top of your own README. Form a small org with authors in the same ecosystem who have traffic, and cross-submit PRs and comment on each other's videos.

**When to move** : the first path has no fixed date. The window is the one or two weeks when the official feature is getting roasted; after that, you're only fixing bugs. If you want to ride the compaction topic in DSH, late November is already too late.

**What if you misread it** : ① If late November it still holds 20+ stars a day, or gets another wave after landing on a bigger official or community list, then DSH's organic traffic for compaction is larger than estimated; ② If someone with no videos and no big account gets 300 stars in two weeks from a similar compaction plugin, then the author's distribution matters less than assumed; ③ If DSH officially changes compaction to compress as it goes, this category of plugin loses its place.

### Data and sources

GitHub: Tranquillitatis-Lab/dsh-live-compact repo page, README, docs/benchmarks.md, commit history, issues #1–#54, daily stars (11-07 to 11-14); star counts of similar repos. npm: download counts for dsh-live-compact, billion-context, dsh-compaction-instant (api.npmjs.org, 11-04 to 11-12). Bilibili: Merkyor's BV1Ud436mEKp (11-03), BV1hBW86wEEj (11-09). DSH official discussion threads #11263, #11773, #11816, #12069. Rankings: dsh.so benchmark leaderboard, awesome-dsh-plugin. Anything marked "unverified" or "inferred" is something we did not verify and are inferring.

[GitHub repo ↗](https://github.com/Tranquillitatis-Lab/dsh-live-compact)

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

Competitors

| Name | Where | Launched | Scale | How it differs from this one |
|---|---|---|---|---|
| @deepseek-ai/dsh-compaction-basic (official compaction) | Big-company features | 2026-08-13 | All DSH users (main repo 280k stars) | Official only calls the model for one big summary at about 80% usage, which is slow and voids the whole prefix cache; live-compact makes small local cuts every turn and still hands the big summary to official as a fallback |
| ranxianglei/billion-context | Open source | 2026-08-06 | 7,266 stars, 9 days on npm with 1.05M downloads | Head of the category, has the model actively prune, works across multiple harnesses; live-compact calls no model at all and focuses on preserving the prefix cache |
| TsFreddie/dsh-compaction-instant | Open source | 2026-08-14 | 17 stars, 159 npm downloads in 9 days | Also model-free, millisecond-level, and readable back, but still compresses one big chunk at the threshold and doesn't discuss caching; no video, no benchmarks, so it never took off |
| Tyan66666/billion-context-dsh | Open source | 2026-08-13 | 169 stars | Model-driven active context pruning |
| mengwEX/dsh-block-ptc | Open source | 2026-10-15 | 103 stars | The agent itself decides when and what to compact |
| dsh-hypercompact / dsh-compaction-pro / -cacheaware / -threshold / dsh-smart-compaction / dsh-compaction-tower and dozens more | Open source | 2026-08~11 | Mostly 1–5 stars | Each takes one angle (summary quality, thresholds, cache repair, deterministic compaction), none had distribution or head-to-head benchmarks |
| Claude Code / Codex have built-in auto-compact | Other | 2025 | Big | Other ecosystems; the author cites them as examples of "hardcoded 80% before compacting" |

Where users come from

- **Verified** **The author's own Bilibili videos** : 11-03 21:55 open-source intro video, 59k views, 4,197 favorites, 2,105 comments; 11-09 22:37 mechanism deep-dive video, 49k views, 2,762 favorites; after the second one, 11-10 / 11-11 stars +17 / +62 and 969 npm downloads that day, the highest of the whole period (Bilibili API; GitHub repo; npm)

- **Verified** **The author's existing local-model audience** : Merkyor has 5,686 followers on Bilibili; his earlier "Qwen3.8 27B cyber therapy" video got 112k / 51k views; issues include local ninfer users; the plugin explicitly splits into cloud and local modes

- **Verified** **Intense engagement in comments and issues** : the author says in the video description he replied to hundreds of comments; in 11 days there were 54 issues/PRs, 13 npm versions, and 196 commits (87 credited to Kimi, 27 to Codex); the share of comments from DSH-driven AI accounts is unverified

- **Verified** **Mutual promotion within a small ecosystem org (Tranquillitatis Lab)** : 11-01 created the org, which co-maintains with T-Auto (dsh-ops, 5,298 stars, dsh-TUI team) and hairyf; T-Auto has 27 commits and several PRs; the dsh-TUI author left a comment under the video

- **Verified** **Community ports and third-party benchmarks** : a Pi port (11-03) and an OpenCode port (11-07, est. cost down 14.3% / 42.1%); the author put the latter at the top of the README

- **Verified** **Plugin directories, leaderboards, aggregator blogs** : dsh.so L5 benchmark leaderboard #8, awesome-dsh-plugin TOP200, dsh-plugin.org, dshfind, dshget, a Yeyu Piaoling blog and others that auto-index it

- Inferred **X, media, and KOL roundup videos** : x-radar and Brave found nothing; I did not verify whether 天才依山尽's 11-12 "Four plugins (Part 2)" covered it. Post-release daily gain was only +11 / +13

## Why dsh-live-compact (Tranquillitatis-Lab/dsh-live-compact) managed to take off

> Channel: GitHub | Research date: 2026-11-13 (Beijing time 11-14) | Data card: repo age 2026-11-03,11 days, 260 stars / 19 forks, last 7 days +147, single-day peak +62 (11-11)

### Bottom line

A DSH compaction plugin built by a Bilibili creator and carried by two of his own videos. The author, Merkyor (GitHub MerkyorLynn), was doing local LLM post-training content on Bilibili with 5,686 followers; his flagship video "Qwen3.8 27B cyber therapy" has 110k views. In early November the DSH community was fed up with official compaction: it waited until 80% usage, took minutes when it ran, wiped the cache, and could even loop forever. He put forward a clear idea at that moment: "Don't wait for 80%. At the end of each turn, make a small local cut, taking tens of milliseconds, with no model call, keeping the prefix cache at 99%." He posted the video that night (59k views, 4,197 favorites, 2,105 comments), then a mechanism explainer on 11-09 (49k views); the peak of +62 stars at 11-11 came after the second video. What set it apart from dozens of similar compaction plugins was a benchmark you could check, the author's heavy replies in comments and issues, and an iteration pace of 54 issues/PRs and 196 commits in 11 days. Still, it remains small in the DSH ecosystem: 5,501 npm downloads in 9 days, versus 1.05M for the head of the category, billion-context, over the same period.

### 1. What it is

A third-party compaction plugin for DeepSeek Harness (DSH, the agent framework DeepSeek open-sourced in 2026-08-13, main repo 280k stars) that can directly replace the official `@deepseek-ai/dsh-compaction-basic`:

- The official approach: only when context reaches about 80% does it call the model for one big summary (the author calls it the "thunderbolt compaction"). It's a long wait, the summary loses detail, and the whole prefix cache is voided.

- How Live Compact works: at the end of each turn it checks, and once context passes 20k tokens it swaps long tool outputs in completed turns, in place, for a "card" (local extraction of excerpts, no model call). The cut point never crosses the previous card, so the prefix already in cache is never rewritten. The originals are backed up locally to `~/.dsh/compaction-spine/shadow` and can be read back by event number. Near the window limit, official still runs one big summary as a fallback.

- It handles cloud APIs (DeepSeek, GLM, where cache hits and cost matter) and local / intranet models (where context level and prefill time matter) separately; the latter is exactly the audience of the author's own local models.

- Install: `dsh plugin --profile web add dsh-live-compact@1.1.7`, plus a domestic mirror the author runs, download.merkyorlynn.com (source: GitHub README).

### 2. Who built it

- **Merkyor / MerkyorLynn** : Bilibili bio "an indie developer who likes post-training", 5,686 followers, 16 videos. Previously built the Lynn desktop agent (GitHub 72 stars), the lynn-engine inference engine (28 stars), and HF models such as Qwen3.8-27B Coder390. Bilibili flagships: "Qwen3.8 27B cyber therapy finale" 112k views (10-08), "Qwen3.8-27B cyber therapy in practice" 51k (09-05) (source: Bilibili search API). His longtime viewers are local-model players, which is why an issue asks "Local Qwen3.8 27b (nvfp4 ninfer) cache hit is only 83%" (#30).

- **Tranquillitatis Lab (Tranquillitatis-Lab)** : a GitHub org created by 2026-11-01 that pulls together repos from several DSH plugin authors: hairyf (dsh-tool-activator, dsh-office-kit), T-Auto / Fengxue (author of dsh-ops, 5,298 stars, a Bilibili creator, also on the dsh-TUI team), and MerkyorLynn. live-compact was first created under MerkyorLynn's personal account (1.0.9 release still points to MerkyorLynn/dsh-live-compact, which GitHub now redirects), and moved into the org around 11-07. T-Auto has 27 commits and several PRs in the repo (hooking into the official settings UI, consolidating upstream dependencies).

- Of 196 commits, 87 are credited to Kimi, 27 to Codex, 1 to Cursor Agent, and 53 are MerkyorLynn's own. The author relies heavily on AI agents for edge work like filling in tests and fixing Windows permissions (source: git log).

### 3. Growth data and timeline

| Time (Beijing) | Event / star count | Source |
|---|---|---|
| 11-01 | Compaction problems kept surfacing in the official discussions: compaction-basic dropped reasoningEffort, so the whole prefix cache missed (#11816); after history exceeded the summary limit, /compact failed 101/101 times and looped (#11773); earlier there was also "compacting at 38K in a 120K window" (#11263), and others | GitHub Discussions |
| 11-01 | Tranquillitatis Lab org created | GitHub API |
| 11-03 18:36 | First commit "Publish Live Compact 1.0.0"; the README was edited 10 times that night, with 150-question / 20-question comparison charts and slides added | git log |
| 11-03 | The author opened discussion #12069 on the official repo, asking that built-in compaction be replaceable by plugins | GitHub Discussions |
| 11-03 21:55 | Bilibili video "DeepSeek Harness real-time self-compaction plugin Live-Compact open-source intro: Cherish your context, stay away from thunderbolt compaction!" (16 min): currently 59k views, 2,166 likes, 4,197 favorites, 2,105 comments | Bilibili API BV1Ud436mEKp |
| 11-03 same day | Someone ported it to Pi (carrotciallo/pi-live-compact, issue #1); T-Auto of dsh-TUI asked about compatibility (#2); dsh-TUI author ccchimneyyy left a comment under the video (48 likes) | GitHub; Bilibili comments |
| 11-04 | npm package published; a dozen or so directory sites auto-indexed it, including the Yeyu Piaoling blog, dsh-plugin.org, dshfind, dshget, dshai.org | npm registry; Exa / Brave search |
| 11-04~06 | npm downloads of 320–840 per day | api.npmjs.org |
| 11-07 | 113 stars (first observation in our database; no day-by-day counts from 11-03~06, the GitHub stargazers API returned no data this time); Haoming9527 started an OpenCode port | github repo_day |
| 11-07~08 | Windows permissions (icacls / SID) crashed the host, bringing 6 high-severity issues at once (#20 #22–#25 #28); the author posted 1.1.0–1.1.2 updates in a single day | GitHub issues |
| 11-09 22:37 | Second Bilibili video "Keep the model's attention as high as possible with the lowest context level possible: design intent and implementation of DSH Live Compact": 49k views, 2,762 favorites, 977 comments. The description says "after the previous video, I replied to hundreds more comments" | Bilibili API BV1hBW86wEEj |
| 11-10 / 11-11 | +17 / +62 stars (the single-day maximum of the whole period); npm 11-10 that day with 969 downloads (the peak) | GitHub repo; npm |
| 11-10 | OpenCode port public (issue #40): Aider 150 / Web-Bench React 20 all pass, estimated cost down 14.3% / 42.1% | Haoming9527/opencode-live-compact |
| 11-12 | The author also built pi-live-compact himself | GitHub |
| 11-13 | v1.1.7, added a GLM-5.3 Flash Web-Bench comparison | docs/benchmarks.md |
| 11-14 | 260 stars / 19 forks; dsh.so "L5 Benchmark" #8 | github repo; dsh.so/zh |

npm: 11-04 to 11-12, 5,501 downloads in total. The two videos together got about 108K views and nearly 7,000 bookmarks, and turned into 260 stars. Bilibili viewers mostly install the plugin directly instead of starring it on GitHub, so the star count is far below its actual exposure in the DSH Chinese community.

### 4. Competitors

The DSH compression niche is very crowded. The official compaction has a poor reputation, and there are dozens of plugins with compaction / compact in their names in the community, mostly in single-digit stars:

| Product | Type | Launched | Scale | How it differs from live-compact |
|---|---|---|---|---|
| @deepseek-ai/dsh-compaction-basic | Official built-in (default) | 2026-08-13 | All DSH users | Waits until about 80% before calling the model for one big summary: slow and costly on cache; live-compact trims small chunks locally every turn, and still hands the big summary to it as a fallback |
| ranxianglei/billion-context | Open source | 2026-08-06 | 7,266 stars, 9 days on npm with 1.05M downloads | The leader in the niche. Has the model actively prune (ACP), works across Claude Code / Codex / OpenCode / pi; live-compact doesn't call the model and focuses on keeping the prefix cache intact |
| Tyan66666/billion-context-dsh | Open source | 2026-08-13 | 169 stars | Model-driven proactive context pruning, same idea as above |
| TsFreddie/dsh-compaction-instant | Open source | 2026-08-14 | 17 stars, 159 npm downloads in 9 days | Closest in idea: zero model calls, millisecond-level, can read back by seq. But it still compresses a big chunk at the trigger threshold, not small cuts every turn, and says nothing about cache; no video, no benchmarks, so it didn't take off |
| mengwEX/dsh-block-ptc | Open source | 2026-10-15 | 103 stars | Lets the agent decide when to compact |
| mrbeandev/dsh-hypercompact | Open source | 2026-09-26 | 3 stars | Deterministic, zero model, compresses by byte budget |
| dsh-compaction-pro / -cacheaware / -threshold / -policy / dsh-smart-compaction / dsh-compaction-tower, etc. | Open source | 2026-08~11 | Mostly 1–5 stars | Some improve summary quality, some tune thresholds, some preserve cache (dsh-compaction-cache claims hit rate fixed from 11.8% to 99.7%); none got distribution |
| bowenliang123/dsh-context | Open source | 2026-08-14 | 5,055 stars | Context visualization: it only "looks" and doesn't compress, complementary to it |
| Claude Code / Codex have built-in auto-compact | Other | 2025 | Big | Other ecosystems; the author's video uses them as counterexamples of "hard-coded 80%" |

The "zero model, local extraction, readable back" idea isn't new: instant and hypercompact did it long ago, so the idea alone doesn't explain why it took off. What it adds is three things: a combination of **small cuts every turn, cut points that never cross old cards, and preserved prefix cache** ; a set of benchmarks against the original; and the Bilibili videos.

### 5. Growth tactics (evidence)

1. **The author's own Bilibili videos (confirmed, main channel)** : 11-03 launch video 59K views, 11-09 explainer video 49K. The audience is mainly the author's existing fans from his local-model work, plus the booming DSH plugin scene (the same-period "eight plugins you must download to use DSH" got 790K views, showing the scene has large traffic). After the second video, 11-10 / 11-11 stars +17 / +62 and 969 npm downloads in a day, the highest of the whole period. This causal link is inferred from timing; we don't have hourly data.

2. **Intense interaction in comments and issues (confirmed)** : the first video got 2,105 comments (comments are 3.5% of views, very high), and the author says he replied to hundreds. The dsh-TUI author also came in to cross-promote. Among the comments are AI accounts that are clearly DSH-plugin-driven, such as "Nureha's whale" and "the big fish that took over its owner's account"; how many of the comments they make up is unverified, so the comment count may be inflated.

3. **Mutual help and organizational backing within the ecosystem (confirmed)** : joined Quiet Sea Lab, and maintains it together with T-Auto (dsh-ops, 5,298 stars, has Bilibili followers, a member of the dsh-TUI team) and hairyf. T-Auto filed a compatibility issue on day one, then submitted PRs, building the link with dsh-TUI.

4. **Secondary spread and trust from community ports (confirmed)** : someone ported it to Pi the same day, and a week later someone ported it to OpenCode and did an independent benchmark. The author put the OpenCode result, "14.3% / 42.1% lower cost", on the front of the README as third-party endorsement.

5. **Directory sites and SEO (confirmed, passive)** : dsh-plugin.org, dshfind, dshget, dshai.org, dshindex, deepseek-harness-plugin.com, zhengdianxue, and the Yeyu Piaoling blog all indexed it automatically; bruc3van/awesome-dsh-plugin TOP200 listed it; dsh.so L5 benchmark ranks it #8. These are automated aggregations in the ecosystem, not something the author pushed.

6. **Feature requests to the official team (confirmed, indirect)** : opened #12069 in the deepseek-harness discussion board, linking to this repo. It has only 2 likes, so it brings little traffic.

7. **X, media, big accounts (not found)** : x-radar shows no "live compact / compress while working / MerkyorLynn" in 30 days; Brave searches for site:x.com and linux.do returned nothing; no media like Zhidongxi, Aikeke, or Cui Tianyi mention it, and no official figure does. Whether the "four must-download plugins (part 2)" by Tiancai Yishan Jin (11-12) included it is unverified; after its release, 11-13 / 11-14 rose only +11 / +13, so no traffic gain is visible.

### 6. Why it took off

- **The pain point is hitting right now** : in early November, the official discussion board kept raising issues like compaction causing full cache misses, /compact loops, and early compaction in small windows; DeepSeek API cache-hit pricing is far below cache-miss pricing, so in long tasks "one compaction wipes the cache" is literally money.

- **Understood in one sentence** : "Don't wait until 80%, compress at any time, tens of milliseconds, cache hits 99%+" (the video description's own words). "Compress while working" and "stay away from the thunderous big compaction" spread much better than the similar "LLM-free lossless compaction engine."

- **The author has an existing audience and shoots his own videos** : 5,686 Bilibili followers, the previous video had 110K views; the two videos together got about 108K views. dsh-compaction-instant, with the same idea, got no distribution at all and stayed at 17 stars.

- **Let benchmarks speak** : on day one it posted 150 Aider tasks and 20 WebUniverse long-task comparison charts, then added a forgetting comparison and a GLM Web-Bench comparison, and cited OpenCode's independent results. The benchmarks are written with restraint, stating which numbers are historical and not DSH results, which makes them credible to technical viewers.

- **Fast responses that won over heavy users** : 11 issues/PRs in 54 days, many of them thousand-word reproduction reports (Windows ACL, data from 32-hour long sessions, freezes across sessions). The author shipped 13 npm versions and 196 commits in a row (most written by Kimi / Codex), replying to every issue, so a group of heavy users stayed to help test.

- **A portable pattern carries its own spread** : the Pi and OpenCode ports each credit back with a backlink in their READMEs, and the author put the third-party benchmark back on his own front page.

### 7. Differentiation

Similar plugins either still "compress a big chunk at the threshold" (official, instant, hypercompact, which only skip the model or write better summaries), or let the model manage it itself (billion-context, block-ptc). live-compact changes the **timing and position of compression** : it cuts only the segment just finished at the end of each turn, cut points never move left past existing cards, so the cached prefix doesn't change a single byte. This keeps the context level low all the time, preserves cache hits, and the original text can still be read back by event number. The difference is real, and it's the basis for the video's "accidentally made something amazing" framing. But the difference alone doesn't explain the growth: similar ideas were built before and went unseen. It took off about 70% from the author's Bilibili distribution and interaction, and about 30% from this difference making the first wave of technical viewers think "worth a try" and willing to file issues.

### 8. What to borrow

- On the days when the platform's official feature had the worst reputation, cut in with a counterintuitive line ("why must it wait until 80% to compress?");

- The author, who already had fans, shot two videos himself, "open-source intro + principle explained," a week apart, the second one specifically answering questions from the first video's comments;

- Released a side-by-side benchmark against the original on day one, with restrained numbers and clearly stated conditions;

- Welcomed other harness ports and put third-party benchmarks back on the front of his own README;

- Had AI agents (Kimi, Codex) handle long issues from heavy users, trading iteration speed for trust;

- Formed a small organization (Quiet Sea Lab) with traffic-holding authors in the ecosystem, cross-filing PRs and promoting each other.

### 9. Still unclear

- No daily star data for 11-03~06: our repo tracking only started at 11-07, and the GitHub stargazers API returned empty this time; how many stars the first video brought can't be separated out.

- The share of AI accounts among Bilibili comments was not verified.

- Whether the Tiancai Yishan Jin "four plugins (part 2)" included it was not verified.

- npm downloads include auto-updates and mirror duplicates, so they can't be converted into real install counts; in addition, the author's own mirror download.merkyorlynn.com diverts some traffic, and that part is not in the npm stats.

- No evidence found of paid promotion or bought comments.

- Stability: users reported hard freezes or no switching at all across 1.1.2–1.1.6 (#34 #49 #53 #54), so whether the reputation can hold up still needs watching.

### Sources

- GitHub: [github.com/Tranquillitatis-Lab/dsh-live-compact](https://github.com/Tranquillitatis-Lab/dsh-live-compact) (README, docs/benchmarks.md, git log, issues #1–#54); organization [github.com/Tranquillitatis-Lab](https://github.com/Tranquillitatis-Lab)

- Bilibili: BV1Ud436mEKp, BV1hBW86wEEj (Merkyor, mid 104669299); BV1GJnZ6MEFw, BV1V5Wd6kEdx (Tiancai Yishan Jin)

- npm: api.npmjs.org/downloads (dsh-live-compact, billion-context, dsh-compaction-instant, dsh-context)

- Official discussions: deepseek-ai/deepseek-harness discussions #12069, #11816, #11773, #11263, #12776

- Competitors: ranxianglei/billion-context, TsFreddie/dsh-compaction-instant, mengwEX/dsh-block-ptc, mrbeandev/dsh-hypercompact, Haoming9527/opencode-live-compact, carrotciallo/pi-live-compact

- Rankings / directories: dsh.so/zh (L5 benchmark #8), dsh-plugin.org, bruc3van/awesome-dsh-plugin TOP200, blog.yeyupiaoling.cn

- Our github repo tables: repo_day, repo_profile (2026-11-07~11-14)

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