DeepSeek Harness (dsh): Version, Updates, Stability (2026)

dsh went quiet for eleven days after launch, then started shipping, and the release channel has moved ahead of npm. npm serves 0.1.1-rc.2; GitHub carries tags for two 0.1.2 alphas that npm has never published. That is a different problem from the one this article described in August, and it changes what “pin your version” has to mean.

npm latest:         0.1.1-rc.2, published 2026-08-21 12:42 UTC
Newest tagged:      dsh-v0.1.2-alpha.2, 2026-08-30 (not on npm)
Versions on npm:    10
Releases / tags:    6 / 6, all prereleases, first on 2026-08-17
Last public commit: 2026-08-30 13:37 UTC
Commits:            14,460 (12,293 at launch)
Branches:           master only
Issue tracker:      disabled, Discussions and Discord only
Discussions:        5,068 threads (snapshot 2026-08-30)
License:            MIT
Official status:    developer preview, breaking changes promised
Entry point:        npx @deepseek-ai/dsh web, Web UI on 127.0.0.1:3080

What Version of DeepSeek Harness Is Current?

It depends where you look, and that is the interesting part. On npm the answer is 0.1.1-rc.2, published 2026-08-21 at 12:42 UTC, with both the latest and next dist-tags resolving to it. On GitHub the newest tagged release is dsh-v0.1.2-alpha.2, cut on 2026-08-30, and it has no npm counterpart.

The full published history on npm is ten versions:

Two versions exist only as GitHub releases: dsh-v0.1.2-alpha.1 (2026-08-27) and dsh-v0.1.2-alpha.2 (2026-08-30). A commit on 2026-08-24 titled “route dsh prerelease dist-tags” shows the team building a prerelease channel system, which is presumably where those alphas are meant to land. But as of 2026-08-30 npm view @deepseek-ai/dsh dist-tags still returns only latest and next, both pointing at 0.1.1-rc.2.

The practical consequence: npx @deepseek-ai/dsh no longer gets you the newest build that exists, and no npx invocation can reach the alphas at all. If you are trying to reproduce a bug someone reported against 0.1.2-alpha.2, npm cannot give it to you.

Has DeepSeek Harness Been Updated Since Launch?

Yes, after an eleven-day gap. The public repository sat at its launch-day commit from 2026-08-13 until 2026-08-24, then resumed. As of 2026-08-30 pushed_at reads 2026-08-30T13:52:13Z and the commit count has gone from 12,293 to 14,460.

What changed structurally, not just numerically:

• Releases and tags now exist. There were zero of each in mid-August. There are now six, starting with dsh-v0.1.0-rc.7 on 2026-08-17, every one flagged as a prerelease. That is a real upstream signal for “this build is a thing,” which the project did not have before.
• A prerelease channel is being built. Two commits on 2026-08-24, “define dsh prerelease channels” and “route dsh prerelease dist-tags,” are infrastructure for shipping alpha and rc streams separately.
• master is still the only branch, and issues are still disabled.

So the earlier read, that you could not tell “quiet because stable” from “quiet because busy,” resolved in favour of busy. The new gap is a different one: the tag stream and the npm stream have come apart, and npm is the one your tooling can actually install from.

How Do You Check dsh’s Status Yourself?

Three commands, no GitHub account needed. Every number in this article moves, so verify rather than trust a snapshot.

Current version and the full publish history:

npm view @deepseek-ai/dsh dist-tags versions time

Whether the public repository has moved at all:

curl -s https://api.github.com/repos/deepseek-ai/deepseek-harness \
  | jq '{pushed_at, has_issues, open_issues_count}'

That returns 2026-08-30T13:52:13Z, false and 0 as of 2026-08-30. Releases and tags are separate calls, and both now return six entries where they returned empty arrays in mid-August:

curl -s https://api.github.com/repos/deepseek-ai/deepseek-harness/releases | jq '.[].tag_name'
curl -s https://api.github.com/repos/deepseek-ai/deepseek-harness/tags | jq '.[].name'

Run both the npm and the GitHub check, not one of them. They disagree right now, and the disagreement is the finding: a tag with no npm publish behind it is a build you can read the source of and cannot install.

Is DeepSeek Harness Production Ready?

No, and DeepSeek says so directly. The README carries its own heading for it:

DeepSeek Harness is in developer preview and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.

Two structural facts back that up rather than contradict it.

Issues are disabled on the repository, so the only official channels are Discussions and Discord. Bug reports and feature requests land in the same feed with no assignee, no label and no status. That is a fine setup for gathering signal and a poor one for tracking a regression you filed.

The plugin architecture points the same way. Everything in dsh is a Cordis plugin, including the UI, which is what makes the ecosystem move fast. It also means a breaking change upstream can land in a surface you did not know you depended on.

Why Did the dsh Repo Go Silent and Then Restart?

Because what got published is the code, not the process around it. The repository was created on 2026-08-13 but arrived carrying its full history: 12,293 commits reaching back to 2026-06-10, whose first commit reads “Initialize repo with README, AGENTS.md, and CLAUDE.md symlink.” Two months of work, all of it visible.

What stayed private is everything around that code. The merge commits reference pull requests #2519, #2520 and #2521 in a separate deepseek-harness organization you cannot open, and master is still the only branch. The eleven-day silence after launch was development continuing in that private org while nothing was pushed out; it ended on 2026-08-24, and the commit count has since climbed to 14,460.

So the code history is unusually transparent and the release cadence is unusually lumpy. Development lands in public in batches, whenever the private side decides to push.

That has one concrete consequence, and it is not the one it looks like. Eleven days of silence did not mean the project had stalled, so the absence of commits is not a health signal here and never was. What it does mean is that you cannot predict when the next batch arrives, and the README promises that batch may break you.

The practical response is to pin. npx @deepseek-ai/dsh@0.1.1-rc.2 costs nothing today and is the only thing standing between you and a promised breaking change.

How Big Is the Community Around dsh?

Large, and still growing faster than the code. GitHub Discussions held 5,068 threads on 2026-08-30, up from 2,713 on 2026-08-17 and roughly 620 three days before that. The repository is at 204,281 stars and 23,614 forks. Star and fork counts move fast enough that any specific number here is a timestamp, not a fact; the direction is the only stable part.

What the volume is made of:

Two weeks ago the useful number pair was thousands of discussion threads against zero commits. That gap has closed on the code side and not on the process side: there is now a real commit stream and a real tag stream, and still no issue tracker to file a regression into. Demand is real, and it has nowhere structured to go.

Note what the ecosystem column implies. When plugins ship faster than the core they plug into, and the core promises breaking changes, the plugins are the first thing that breaks. That is an argument for keeping your dsh install thin rather than for avoiding it.

Does dsh Show You What Each Task Costs?

Not today. There is no per-turn token or cost display in the shipped build. One of the more upvoted Ideas threads, filed 2026-08-14, is a request to add exactly that, and a community cost-tracker plugin appeared on launch day to fill the gap.

The timing makes this sharper than it looks. DeepSeek moved its whole V4 family to peak and off-peak pricing on 2026-08-16, so the same agent loop now costs one of two prices depending on the hour.

Cache hits on deepseek/deepseek-v4-pro-0813 went up 12.1x at peak, and deepseek/deepseek-v4-flash-0731 5x. An agent harness that resends the same system prompt and file contents every turn is precisely the workload that change targets.

Until the harness surfaces it, your options are the provider dashboard, your own proxy logs, or one of the community plugins built to show cost per task and peak-hour status. None of those ships with dsh.

When Is dsh Safe to Use Today?

Safe to explore, not safe to depend on. The split is cleaner than the developer-preview label suggests, because it comes down to what breaks when the next version lands.

The dividing line is whether an upgrade that breaks your setup costs you an afternoon or a deadline. DeepSeek has told you in capital letters which of those it is planning.

What Should You Do Before Wiring dsh Into Your Workflow?

Pin the version, keep the model endpoint swappable, and do not put it on the critical path yet. The first two cost nothing. The third is the actual decision.

A concrete checklist:

1. Pin it. Run npx @deepseek-ai/dsh@0.1.1-rc.2, never bare dsh. The version moved four times in the eight days after launch, so bare dsh is a moving target now in a way it was not in mid-August.
2. Point it at a base URL you control, not a hardcoded vendor default.
3. Keep a second harness working for anything with a deadline.
4. Watch both npm and the GitHub releases feed. That advice used to be “npm only.” It is wrong now: the two 0.1.2 alphas were tagged on GitHub with no npm publish behind them, so npm alone will show you a version behind.
5. Reread the README after any upgrade. Breaking changes are announced there and nowhere else.

Point 2 is the one people skip. dsh reads DEEPSEEK_BASE_URL and DEEPSEEK_API_KEY for its built-in route, and the custom-provider form accepts any OpenAI-compatible endpoint, so the model layer is genuinely swappable if you set it up that way from the start.

On ofox that means deepseek/deepseek-v4-flash-0731 and deepseek/deepseek-v4-pro-0813 behind one key. The value is not the gateway itself. It is that on the day dsh breaks, or the day pricing moves again, the model string is the only thing you edit.

The step-by-step for both routes is in the dsh custom provider setup guide. For how dsh sits against Claude Code, Codex and the rest on extensibility and model lock-in, see the coding agent harness comparison.

References

• deepseek-ai/deepseek-harness on GitHub
• dsh README, developer preview notice
• DeepSeek Harness Discussions
• dsh releases and tags
• @deepseek-ai/dsh on the npm registry
• Cordis, the plugin runtime dsh is built on
• DeepSeek API pricing
• DeepSeek change log