DEV Community

Owen
Owen

Posted on Originally published at ofox.ai

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

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

DeepSeek Harness remains at version 0.1.0-rc.6, its launch build from August 13, 2026. The public repository has received no commits since that date, with no releases, tags, or active issue tracking. Pin your version before building anything with this tool.

Current Status Summary

The tool launched with six published versions between August 10-13, 2026, all available on npm under both latest and next dist-tags. The only branch in the repository is master, and discussions have grown to 2,713+ threads while the codebase remains static since launch.

Key facts:

  • Current version: 0.1.0-rc.6 (August 13, 2026 12:35 UTC)
  • Total versions published: 6
  • Last public commit: August 13, 2026 11:38 UTC
  • Releases/tags: 0 / 0
  • License: MIT
  • Official status: Developer preview with promised breaking changes
  • Entry point: npx @deepseek-ai/dsh web on 127.0.0.1:3080

Version History

Version Published (UTC)
0.0.1-rc.1 2026-08-10 19:41
0.0.1-rc.2 2026-08-11 15:24
0.0.1-rc.5 2026-08-12 22:36
0.1.0-rc.2 2026-08-13 09:48
0.1.0-rc.3 2026-08-13 11:16
0.1.0-rc.6 2026-08-13 12:35

Every release predates or matches the public launch date.

Checking Status Independently

Three commands verify current state without requiring a GitHub account:

npm view @deepseek-ai/dsh dist-tags versions time
Enter fullscreen mode Exit fullscreen mode
curl -s https://api.github.com/repos/deepseek-ai/deepseek-harness \
  | jq '{pushed_at, has_issues, open_issues_count}'
Enter fullscreen mode Exit fullscreen mode
curl -s https://api.github.com/repos/deepseek-ai/deepseek-harness/releases
curl -s https://api.github.com/repos/deepseek-ai/deepseek-harness/tags
Enter fullscreen mode Exit fullscreen mode

As of August 17, these commands show no movement since the initial publication.

Production Readiness

The project explicitly states it remains "in developer preview and is iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES." This is not merely marketing language — structural choices reinforce it.

The repository disables issues entirely, routing all communication through Discussions and Discord. That works for signal gathering but provides no mechanism for tracking a regression you filed. The plugin architecture, built on Cordis, means breaking changes upstream can land in a surface you did not know you depended on.

The Paradox of Transparency and Opacity

The repository was created August 13 but arrived with complete history: 12,293 commits reaching back to June 10, 2026. The first commit reads "Initialize repo with README, AGENTS.md, and CLAUDE.md symlink." That is two months of visible development history.

What stayed private is the surrounding process. Merge commits reference pull requests (#2519, #2520, #2521) in a separate, non-public deepseek-harness organization. No branches exist beyond master, and nothing has shipped since launch day.

This creates an unusual situation: the code's development path is transparent while its future direction is opaque. "Iterating rapidly" is likely accurate but unobservable from outside. For consumers, that means you cannot distinguish "quiet because stable" from "quiet because actively developing."

Community Size and Patterns

GitHub Discussions grew from roughly 620 threads three days earlier to 2,713 by August 17. Common themes:

Theme Requests Implication
Native client instead of browser Real CLI, desktop app, editor extensions Web-UI-first design is the most contested choice
Install failures Windows, Arch Linux, Termux reports Non-macOS platforms are unverified
Third-party plugins Marketplaces, knowledge bases, UI skins Ecosystem outrunning core
Cost visibility Per-turn tokens, peak-hour indicators Shipped build lacks this
Sandbox and permissions Multiple independent reports Preview disclaimer applies

The ecosystem is expanding faster than the core, which carries a consequence: when plugins ship faster than the foundation they depend on, plugins are the first casualty when breaking changes arrive.

Cost Transparency Gap

The shipped build provides no per-turn token count or cost display. A highly upvoted Ideas thread filed August 14 requests exactly this. A community cost-tracker plugin appeared on launch day to fill the gap.

The timing matters. DeepSeek shifted its entire V4 family to peak and off-peak pricing on August 16, 2026. Cache hits on deepseek/deepseek-v4-pro increased 12.1x at peak; deepseek/deepseek-v4-flash rose 5x. An agent harness that resends identical system prompts and file contents every turn is precisely the workload that pricing change targets.

Until the harness surfaces cost data, you consult provider dashboards, proxy logs, or community plugins. None ship with dsh.

When dsh Is Safe to Use

The practical distinction is what breaks when the next version appears:

Safe to explore:

  • Evaluating the plugin model on scratch repositories
  • Comparing DeepSeek models against your own tasks
  • Building or testing dsh-plugin projects
  • Local experiments you can redo in an afternoon

Hold off on:

  • Anything with a delivery date attached
  • Shared team configurations others will inherit
  • CI or unattended scheduled jobs
  • Work where an unexpected sandbox change would be costly

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 it is planning.

Recommended Setup Approach

Before integrating dsh into production workflows:

  1. Pin the version: Use npx @deepseek-ai/dsh@0.1.0-rc.6 exclusively, never bare dsh
  2. Control the endpoint: Configure DEEPSEEK_BASE_URL and DEEPSEEK_API_KEY pointing to a base URL you manage, not a hardcoded vendor default
  3. Maintain a backup harness: Keep a second working harness for deadline-critical work
  4. Monitor npm, not GitHub: The next version appears on npm first
  5. Review release notes: Reread the README after any upgrade, since breaking changes are announced there and nowhere else

Point 2 carries particular weight. dsh reads environment variables for its built-in route, and the custom-provider form accepts any OpenAI-compatible endpoint. That swappability only works if you configure it from the start. On ofox, that means deepseek/deepseek-v4-flash and deepseek/deepseek-v4-pro behind a single key, so on the day dsh breaks or pricing moves again, the model string is the only thing you edit.

Frequently Asked Questions

What is the current version?
0.1.0-rc.6, published August 13, 2026. Both npm's latest and next tags point to it.

Are there GitHub releases or tags?
No. Zero releases and zero tags exist. Version history lives entirely on npm.

Why are issues disabled?
The maintainers route all communication through GitHub Discussions and Discord instead, eliminating formal issue tracking.

Is dsh a fork?
Neither fork nor mirror. The repository was created August 13 with full commit history from June 10, but the surrounding development process remains private.

Does it run on Windows or Linux?
It ships as a Node package, so it starts anywhere Node runs. However, community reports of install failures on Windows, Arch Linux, and Termux are frequent enough to treat non-macOS platforms as unverified.

Can I pin to a known-good version?
Yes: npx @deepseek-ai/dsh@0.1.0-rc.6. Because no upstream tags exist, the npm version string is your only safeguard against promised breaking changes.

Is dsh free?
The harness itself is MIT licensed and free. Inference costs depend on your provider's rates. DeepSeek's pricing moved to a peak/off-peak structure on August 16, 2026.

Key References


Originally published on ofox.ai/blog.

Top comments (0)