Connect DeepSeek Harness v0.1.1-rc.2 to the DeepSeek API

DeepSeek Harness v0.1.1-rc.2 is the latest official release as of 26 August 2026. It was published on 21 August, five hours after rc.1. Despite the shorthand people use online, this is not a stable v0.1.1 release: the rc suffix matters, and the project remains a developer preview with breaking changes expected.

Official releases: rc.2 · rc.1

What changed in rc.1

DeepSeek V4 Flash Vision

The DeepSeek adapter added the experimental multimodal model DeepSeek-V4-Flash-Vision-Exp. This is more than a model-name addition: Harness must admit an image, store its canonical representation, project it into a model request and retain enough information for replay.

Bubblewrap sandbox escape fix

Confined Linux processes could potentially use /proc/<pid>/root to reach outside the intended Bubblewrap view. rc.1 isolates the PID namespace to close that route. This is the most important operational reason to move off older release candidates when running untrusted tools on Linux.

The fix does not make arbitrary plugins safe. A plugin runs with the privileges granted by its composition, while the sandbox governs selected child processes. Review third-party code and keep permission prompts enabled.

Interaction and Web UI improvements

  • Editing text before an @ reference no longer risks a broken composer layout.
  • ask_user_question accepts multiline answers, wraps long text and uses Shift+Enter for a newline.
  • Wide Markdown tables respond better to the conversation width.
  • Cache-hit values in the 99.x% range retain useful precision.
  • Navigation between a parent session and subagent sessions is clearer.

What rc.2 adds

rc.2 concentrates on the image pipeline:

  • The DeepSeek adapter prefers the Files API when uploading images.
  • Uploaded files can be reused instead of retransmitted for every request.
  • Images are automatically resized and converted to a format suitable for the selected model.
  • If Files API resolution fails, the adapter has a fallback path rather than making the whole request depend on one upload route.

The practical benefit is lower repeated-upload overhead and fewer failures caused by oversized or unsupported image inputs.

Upgrade

For an npm launch, request the current package explicitly:

npx @deepseek-ai/dsh@latest web

Check the resolved version before diagnosing a behavior difference:

npx @deepseek-ai/dsh@latest --version

Back up ~/.dsh before moving an important installation between release candidates. The project does not promise compatibility for every preview-era configuration or session artifact.

Should you upgrade?

SituationRecommendation
You use images or the Vision modelUpgrade; rc.1/rc.2 are the relevant baseline
You run Linux tool calls in BubblewrapUpgrade for the /proc/<pid>/root fix
You depend on a production workflowTest a copy of the DSH home and workspace first
You only evaluate text tasksUpgrade is still sensible, but the visible feature gain is smaller

What this release does not mean

It does not make Harness stable, turn Agent Teams into a public default, or certify community plugins. v0.1.1-rc.2 is a fast-moving preview release with a better vision path and a meaningful sandbox fix.

Related guides

FAQ

Is DeepSeek Harness v0.1.1 stable?

No. The latest official tag is dsh-v0.1.1-rc.2; it is a release candidate in a developer-preview project.

Does rc.2 include the rc.1 security fix?

Yes. rc.2 follows rc.1 and includes the Bubblewrap PID-namespace fix plus the newer image-pipeline work.