deepseek-ai/deepseek-harness, published as @deepseek-ai/dsh. It is not an unscoped Python package or a thin DeepSeek API client.Runtime or API wrapper?
Profiles, Agent loop, tools, policy, sandbox, Sessions, model adapters, MCP, and UI.
Request formatting, streaming, reasoning fields, retries, or a model-specific client.
Use owner, package scope, executable, and product boundary—not the repository name.
Current official coordinates.
- Repository
deepseek-ai/deepseek-harness- Package
@deepseek-ai/dsh- GitHub release
dsh-v0.1.0-rc.8- npm tags
latest rc.7 · next rc.8- Commit
141eb6f- Handbook
- Independent SandBase project
Read metadata before executing code.
npm view @deepseek-ai/dsh \
name version dist-tags repository.url homepage --jsonThen prove what the shell selected.
git remote get-url origin
git rev-parse HEAD
command -v dsh
dsh --version
Do not infer ownership from an unscoped package, repository title, executable name, copied installation command, or a fork README.
Common questions.
deepseek-ai/deepseek-harness
@deepseek-ai/dsh
A matching PyPI name does not prove official ownership.
No. Model providers are adapters inside the runtime.
Send work to the right owner.
Bugs, proposals, plugin showcases, and upstream fixes.
Operator guides, corrections, navigation, and reproducible evidence.
Cite the commit and state that it has not shipped upstream.
Primary evidence.
Keep the provenance record.
The canonical guide includes install-safe metadata checks, source checkout verification, exact-version execution, contribution routing, and a reusable incident record.
Read and star on GitHub