DeepSeek Harness, 39 days later

Five weeks ago we wrote up DeepSeek Harness and called it the agent layer nobody owns yet. We audited the adoption numbers, found every circulating figure wrong, printed the commands so you could check us, and closed with five things to watch. It is 39 days later. We re-ran the same commands. Here is what the numbers say now, and which of the five predictions actually landed.

The same three commands, 39 days on

The star curve is still steep and still the least informative thing on the table. The number that matters is the fork count, which grew faster in relative terms than stars did, and the plugin repo count, which came close to doubling. People are not just bookmarking this, they are building on it.

Prediction one: do the Issues reopen?

No. Thirty-nine days and sixty-six thousand stars later, has_issues is still false. Discussions remain the only channel.

We said this was the single clearest signal of whether dsh is a community project or a published artifact with a comment box, and by that test the answer has not changed. But the picture is more interesting than the flag suggests, because contribution is clearly happening anyway. The release notes for the newest build credit named outside contributors by handle, several per release, for things like animation timing, image retry reliability and plugin installer layout. So people are landing changes through pull requests and Discussions while the bug tracker stays shut.

That is a deliberate choice about support surface rather than about openness, and five weeks of evidence says it is holding. If you are betting a product on dsh, plan on having no public issue history to search when something breaks.

Prediction two: does the plugin API stabilise?

Half. We said to watch for the release cadence to slow and a 0.2 to appear with a migration note. The 0.2 arrived, today in fact: dsh-v0.2.0-rc.1, published 28 September, and it is sitting on the next npm tag while latest serves 0.1.7-rc.2.

The cadence did the opposite of slowing. Six releases landed in the eleven days to 28 September, and the version count went from 8 to 28. The release notes are bilingual and substantial, but they read as a changelog of user-facing fixes rather than a plugin migration guide.

So the advice from August stands unchanged: treat plugin development against dsh as prototyping. A 0.2 release candidate arriving on a tag that most installs will not pull, during the fastest release month the project has had, is not the stability signal we were looking for.

Prediction three: does anyone else adopt Cordis?

No, and this is the one that matters. We said that if a competing harness could load a dsh plugin, the layer starts to standardise and DeepSeek's bet pays off, and if not, this is one lab's very good internal tool with an MIT license attached.

What grew instead is an ecosystem entirely inside DeepSeek's orbit. A web plugin aggregator sits at 8,100 stars. The first vision plugin is at 4,000. A desktop shell, started the same day as dsh itself, is at 29,300. There is a plugin directory at 17,200, and a widget that watches your DeepSeek account balance from the corner of the screen has 3,300. Cordis itself, the underlying meta-framework, has 8,800 and remains an independent project.

Every one of those is a dsh plugin or a dsh surface. None of them is a rival harness that speaks the protocol. Sixteen thousand plugin repositories is a genuine ecosystem by any measure, but it is a deep ecosystem rather than a wide standard, and the distinction decides whether your plugin work is portable. Right now it is not.

Predictions four and five, briefly

On independent benchmarks using Minimal mode, there is a development worth noting even though it is not the one we asked for. DeepSeek-V4.1-Flash shipped in September with its agentic benchmark table run in the Minimal mode of DeepSeek Harness at a 1M-token context. That is the preset being used for exactly the cross-model comparison we hoped for, but it is DeepSeek evaluating DeepSeek inside DeepSeek's harness, which is the opposite of independent. The open leaderboard we were waiting for has not appeared.

On the pricing split, we have seen nothing that changes the August reading, so treat that one as still open.

What the scorecard actually says

One of five predictions landed, and only halfway. That is worth stating plainly, because the useful thing about writing down what to watch is that you later find out what you got wrong.

What we underestimated is how fast a plugin ecosystem can form around a closed support surface. We framed disabled Issues as the thing that would cap community growth, and it did not: sixteen thousand plugin repos and ten thousand new forks happened with the bug tracker shut the entire time. The friction we expected simply did not materialise, which says something about how little a public issue tracker matters to people writing plugins against a stable-enough API.

What we called correctly is the portability risk. Five weeks of extraordinary growth has produced no evidence that anything outside DeepSeek can load a dsh plugin. The bet the August piece described, that dsh becomes the standard harness layer rather than one lab's runtime, is no closer to paying off than it was, and the ecosystem getting bigger does not move it. If anything a larger captive ecosystem makes a competing implementation less likely, not more.

What to watch in the next 90 days

The three questions still worth tracking, narrowed by what we now know.

Whether 0.2.0 reaches latest with an actual plugin migration guide, rather than sitting on next indefinitely. That is the moment plugin authors get to stop prototyping.

Whether any non-DeepSeek harness ships Cordis plugin loading. This remains the single fact that would change what dsh is, and it has not happened in the five weeks when the incentive was strongest.

Whether a third party publishes a Minimal-mode leaderboard. DeepSeek has now demonstrated the preset works for cross-model evaluation by using it on its own model card, which removes the last technical excuse for someone else to run it properly.

Sources and further reading

• DeepSeek Harness on GitHub
• The release history, including 0.2.0-rc.1
• The community plugin directory
• dsh-desktop, the third-party desktop shell
• Cordis, the underlying plugin meta-framework

Ten minutes: run the same three API queries we published in August against whatever project you are currently betting on, and write down five things you expect to be true in five weeks. The value is not in being right. It is in having something specific to check when the star count has doubled and everyone has decided the question is settled.

Tested on: not independently tested. All figures are read directly from the public GitHub, npm and GitHub Search APIs on the dates shown, using the same queries published in our August article so the comparison is like for like. Star and fork counts move continuously; the 2026-09-28 column was captured in a single pass on that date.
Date checked: 2026-09-28