DSH v0.1.2-alpha.3 Improves Long Sessions and Queued Images
DeepSeek Harness v0.1.2-alpha.3 improves long-session navigation and rendering, fixes queued image delivery, and removes the optional SQLite Session backend. The alpha remains separate from npm's default RC release.
DeepSeek published DSH v0.1.2-alpha.3 at 16:03 UTC on August 31, 2026, which was 00:03 on September 1 in Beijing. The release is the newest DeepSeek Harness preview, but it is still explicitly marked as a pre-release and developer preview.
That status matters because DSH is moving on two public release lines. The npm latest and next tags still point to 0.1.1-rc.2, while the alpha tag points to 0.1.2-alpha.3. Running the unversioned quick-start command therefore does not install alpha.3.
# npm latest: 0.1.1-rc.2
npx @deepseek-ai/dsh web
# exact alpha.3 preview
npx @deepseek-ai/dsh@0.1.2-alpha.3 web
Pinning the exact version is the clearer choice for a preview evaluation. It makes the test reproducible and avoids silently following a later alpha tag.
Long conversations get easier to navigate
The largest user-facing change is in long sessions. The right-hand conversation navigator can now preview and jump to every paginated turn, including turns that have not yet loaded into the current view. DeepSeek also reports lower memory use and smoother syntax highlighting while rendering long conversations.
Those changes target two different failure modes. Navigation makes an older turn reachable without manually loading page after page, while the rendering work reduces the cost of keeping a large technical conversation usable once it is on screen.
The release also refines multilingual permission labels. This is presentation work rather than a change to the underlying permission boundary.
Queued images now survive active turns
Alpha.3 fixes image handling when a user adds or queues a message while a conversation is already running. The image should now appear immediately and be delivered reliably after the active turn. Follow-up messages sent to continuable subagents use the same repaired path.
The read_image tool can also identify image attachments from their file contents when the attachment path has no filename extension. That closes a practical gap between uploaded attachment paths and tools that previously relied on .png, .jpg, or another suffix.
Smaller interaction and reliability fixes
The command menu now completes the highlighted slash command when the user presses Tab. A backend stall should no longer be misclassified as a lost network connection, and the active Schedule list in a conversation header should stay aligned inside narrow viewports.
These are narrower changes than the new subagent controls and interface restructuring introduced earlier in the 0.1.2 alpha line. Alpha.3 is mostly a usability and reliability pass on top of that work.
SQLite Session persistence was removed
The operational warning is the removal of the optional SQLite Session persistence backend. DeepSeek says existing content is not deleted, but users who need to export it must use an older DSH version. Anyone testing alpha.3 against an existing Harness home should check the active persistence backend before upgrading.
The default JSONL session path and the Python SDK version are separate concerns. PyPI still lists deepseek-harness-sdk 0.1.1rc1; the Python package did not move to alpha.3 with the npm CLI and source tag.
This is a Harness update, not a model release
This is not a DeepSeek model or API release. DSH v0.1.2-alpha.3 does not announce a new model, model ID, API endpoint, context window, or token price. It updates the open-source agent harness around those models. Model and API release claims should remain separate from this tooling release.
The source tree still requires Node.js ^22.19.0 or >=24.0.0. DeepSeek also continues to warn that Harness is not security-audited or production-ready and that its sandbox, approvals, and permissions do not guarantee isolation. The new release is most appropriate for deliberate preview testing, especially when its long-session or image fixes solve a current problem.