DeepSeek Harness v0.1.3-alpha.2: Subagent & Session Fixes

Author
DeepSeekAgent.io Editorial Team
Published
Updated

DeepSeek published the dsh-v0.1.3-alpha.2 GitHub pre-release on September 7, 2026 and released @deepseek-ai/[email protected] to npm's alpha channel. It builds on alpha.1's Session v2 and general-file work, with a focus on continuable-subagent controls, long-session performance, Web reconnection, and Windows runtime fixes.

Installation channels still matter. As of September 8, npm's latest and next tags remain on 0.1.2-rc.1. Only an explicit alpha or exact-version selection installs 0.1.3-alpha.2.

The short version

Alpha.2 addresses alpha.1's acknowledged long-session performance regression and adds substantially better control over continuable subagents. It is still a pre-release, however, and includes compatibility changes to default tools, Persona configuration, and subprocess handles. Regular users should remain on rc.1; developers evaluating the new behavior should pin the version, back up Sessions, and test in an isolated environment.

Release and installation status

CheckStatus
GitHub Releasev0.1.3-alpha.2, marked Pre-release
GitHub tagdsh-v0.1.3-alpha.2 at 82a5fd6
PublishedSeptember 7, 2026 at 13:59 UTC
Exact npm version@deepseek-ai/[email protected] is published
npm alpha0.1.3-alpha.2
npm latest / next0.1.2-rc.1
PyPI SDKdeepseek-harness-sdk==0.1.2rc1

The unpinned command still resolves to rc.1:

npx @deepseek-ai/dsh web

Use the alpha channel or an exact version only when intentionally testing the pre-release:

npx @deepseek-ai/[email protected] web

Full message control for continuable subagents

Continuable subagents now support queued messages, editing, deletion, Steer for one or all queued messages, and Stop. This moves them beyond a start-and-wait interaction toward collaborators that can be redirected while they run.

That is useful for extended research, code review, and multi-stage execution, but it also makes UI, queue state, and recovery consistency more important. A queued message displays a Sending state and cannot be edited, deleted, or steered until delivery completes, preventing an unsent edit from looking effective.

Better long-session performance and references

The official notes say this release reduces lag and memory use when opening, resuming, and continuing long conversations, following the performance regression explicitly acknowledged in alpha.1. When a long conversation is referenced, the model can also read material omitted from the preview as needed.

This does not remove every migration risk. Session v2, lifecycle-scoped SessionHandles, asynchronous agentLoop.create(), and Session locking remain important compatibility boundaries inherited from alpha.1. Back up $DSH_HOME and test opening, continuing, compacting, and recovering real historical Sessions before adopting the release.

Web and local-runtime fixes

The release fixes or improves several everyday paths:

  • Web connections recover automatically after a disconnection.
  • Chat scrolls to the bottom again after a message is sent or the window is resized.
  • A possible Windows Python SDK runtime startup crash is fixed.
  • Process cleanup improves on Windows and supported Linux environments, reducing leftover processes after a task stops.
  • An Open in action in the Web header opens a Workspace in an installed editor, IDE, terminal, or file manager.
  • PTC mode can expand commands and their output.

The latest public Python SDK on PyPI remains 0.1.2rc1. A fix described for the 0.1.3 source and npm line should not be presented as something every default PyPI user already receives.

Model, feedback, and interface changes

pi-ai moves to 0.85.1 for newer-model support. Settings labels, switches, and plugin-status indicators receive a more consistent treatment. Feedback can also be submitted without continuing the conversation, with relevant conversation content attached to the report.

That feedback path may include conversation context. Teams should confirm the submitted scope, internal policy, and whether related conversation material may be sent to an external service before using it.

Compatibility checks before upgrading

Alpha.2 includes three changes that can affect plugins and automation:

  1. SDK, Headless, and ACP use read, write, and edit for file editing by default; Web minimal and sdk-minimal are unchanged.
  2. Custom Persona configuration is split into a prefix and suffix, so old configuration and plugins using related constants may require updates.
  3. Ordinary subprocess handles no longer expose pid; Terminal handles are unaffected.

An upgrade check should cover more than whether Web boots. Test custom Profiles, plugin dependencies, Persona patches, subprocess management, historical Sessions, and headless tasks.

How to read early community reports

Community discussions posted on September 8 report Session v2 migration and third-party HTTP-channel injection problems. The reports use post-alpha.2 master commits or depend on specific third-party providers and plugins. They are useful risk signals, not official confirmation that the exact dsh-v0.1.3-alpha.2 tag necessarily has the same defects.

Environments that depend on historical Sessions, third-party Connection channels, or local OpenAI-compatible providers should reproduce their own configuration before upgrading.

Who should upgrade

ScenarioRecommendation
Regular npm userStay on the default 0.1.2-rc.1
Needs subagent queues and SteerTest exact version 0.1.3-alpha.2 in isolation
Hit alpha.1's long-session regressionEvaluate alpha.2 after backup; do not overwrite the only environment
Uses the Windows Python SDKConfirm the installed channel and version before testing the startup fix
Authors plugins or ProfilesFocus on default tools, Persona, and subprocess-handle changes

Related reading

Frequently asked questions

Can v0.1.3-alpha.2 be installed from npm?

Yes. As of September 8, 2026, the exact version is published and npm's alpha tag points to it. latest and next remain on 0.1.2-rc.1.

Does an unpinned install upgrade to alpha.2?

No. npx @deepseek-ai/dsh web currently resolves through latest, which is rc.1. Only @alpha or exact @0.1.3-alpha.2 selection enters the alpha.2 line.

Has alpha.2 fixed every Session v2 problem?

The official notes describe improved long-session loading, resumption, and memory use, but do not claim that every historical-session and third-party-provider migration combination has been validated. Back up important Sessions and test against a copy.