DeepSeek Harness Releases v0.2.2-alpha.3 Pre-release Version

Smart Things reported on November 4 that yesterday, DeepSeek Harness quietly updated its GitHub repository, releasing a v0.2.2-alpha.3 pre-release version.

This update expanded around code review, automated task management, SDK, and the plugin ecosystem.

Interestingly, just yesterday DeepSeek experienced partial outages of its web and API services.

According to DeepSeek's previous "glorious track record" of having an update every time it crashes, although there is no evidence directly linking the two events, netizens still speculated one after another about whether a new model was coming.

Sure enough, the law of "every crash brings an update" is still too reliable.

In this update, DeepSeek Harness added an experimental unified entry point for "change requests and worktrees," allowing users to view local code and PR diffs through a file tree and hand selected code snippets and comments to the Agent for processing.

Automated tasks were also upgraded to host-level unified management, supporting manual creation, direct editing, and task execution in both new and old sessions.

In addition, DeepSeek Harness added a multi-file modification tool, apply_patch, and further improved the capabilities of the Python and TypeScript SDKs.

1. More convenient code review: DeepSeek Harness optimizes the development collaboration process

In the real software development process, code generation is only one part; developers also need to inspect changes, handle code diffs, follow up on review comments, and ensure the code can be merged smoothly.

In this update, DeepSeek Harness launched an experimental unified entry point for "change requests and worktrees," integrating code review and Git Worktree-related operations together.

Specifically, developers can browse local code and PR diffs through a searchable file tree and view changes file by file. After finding issues, they can also select specific code snippets, add comments, and hand this information to the Agent for processing.

At the same time, this feature also supports tracking inspection and review status.

This allows developers to view code changes, raise modification suggestions, and let the Agent continue processing based on feedback in the same entry point, reducing the need to switch between different tools.

Another noteworthy update is the addition of the apply_patch tool.

In the past, when an Agent modified multiple files, it often needed to perform file reading, editing, and other operations separately. The newly added apply_patch supports users in completing the addition, editing, deletion, and moving of multiple files in a single call, and the modification results can also be viewed in Review.

However, this tool is disabled by default and needs to be actively enabled by the user.

These two updates mean that after an Agent modifies code, developers can directly inspect changes and raise comments in DeepSeek Harness, then have the Agent continue adjusting, making the connection between code modification and review smoother.

2. Unified management of automated tasks lets Agents continue working across sessions

In addition to code development, DeepSeek Harness also made fairly noticeable adjustments to automated tasks in this update.

Previously, automated tasks were more associated with the execution process in specific sessions. In this update, DeepSeek Harness unified automated task management at the Host layer, supporting users in manually creating tasks or directly editing existing tasks, and users can also create tasks in sessions through prefilled prompts.

In terms of execution, users can choose to reuse an existing session or create a new session to execute tasks. In addition, when editing tasks, drafts and conflict review are supported to prevent content from being accidentally overwritten during editing.

This means that automated task management is no longer limited to a single session; users can manage from a higher level and arrange for Agents to execute work in different sessions as needed.

For development tasks that need repeated handling, this change can reduce the operations of repeatedly entering instructions and manually starting tasks.

3. SDK and plugins continue to expand, mobile experience optimized

The expansion of SDK and plugin capabilities mainly concerns how developers build their own Agent applications based on DeepSeek Harness.

In this update, DSH further improved the TypeScript and Python SDKs, filling in native command and plugin management capabilities, including three functions: plan review, goal continuation, and operation approval.

Among them, plan review allows users to approve, modify, or cancel plans submitted by the Agent; goal continuation supports saving long-term goals and continuing to advance them after a session restarts, resumes, or forks; operation approval has the SDK receive approval requests from the Agent and then hand them to the integrated application to decide whether to approve.

On the plugin side, this update allows plugins to provide navigation icons for their own settings pages, and also supports dynamically providing API request headers based on session identity and request purpose.

In addition, when plugin installation fails, more complete diagnostic information and a Host log entry are now provided to help developers locate problems.

Beyond the developer ecosystem, DSH is also continuously improving the mobile experience, optimizing sidebar navigation and narrow-screen layout. However, the "mobile" here is not a standalone app, but access to the Web GUI through a mobile browser.

In terms of interaction and file management, DSH also made adjustments: it unified the right-click menus and keyboard menus for sessions, workspaces, and the sidebar, and opened up developer extension capabilities; file preview added a persistent directory tree to help users quickly locate and switch files.

In addition to the above features, DSH this time also optimized the display of Trajectory call records, supporting independent collapsing of sub-calls, while improving the performance of streaming assembly for long replies.

In terms of bug fixes, the new version resolved issues including historical sessions failing to open, session search failures, chat history loading getting stuck, and some models continuously failing requests when the number of images exceeds the request limit.

In addition, the new version also adjusted the file modification security policy: the pre-modification read check and external modification protection are disabled by default, while sandbox restrictions and the Review mechanism remain.

Final thoughts:

It should be noted that v0.2.2-alpha.3 is still a pre-release version, and the official corresponding desktop installation package has not yet been provided, so existing desktop App users cannot directly upgrade to this version for the time being.

Overall, this update shows that DSH is filling in multiple links needed for Agents to participate in software development, from code modification and diff review to task scheduling and plugin expansion, gradually forming a more complete development workflow.