DeepSeek Harness v0.2.1-alpha.2 Preliminary Release Issued with Reasoning Translation and Git Worktrees Features
Read more
Pandaily
pandaily.com

DeepSeek Harness v0.2.1-alpha.2 Preliminary Release Issued with Reasoning Translation and Git Worktrees Features

DeepSeek has released a preliminary version of its open agent runtime—DeepSeek Harness v0.2.1-alpha.2. This build was posted on the official project page on GitHub on the evening of October 9th Beijing time, six days after the release of v0.2.1-alpha.1, and remains in preliminary release status rather than being a stable version.

The main new feature is an experimental plugin that performs machine translation of model reasoning results. Users can activate this function on the Plugins page and select a translation provider from Bing, Google, or the paid DeepSeek Flash. A second experimental plugin adds support for Git Worktrees, allowing the agent to create and switch to a separate Git checkout so that parallel work does not conflict within a single working copy. Additionally, the experimental plugins have received new session state APIs for saving and reading custom records.

Several changes were made for developers using the runtime. A new working_directory tool allows the agent to read and modify the session's working directory, and the TypeScript and Python SDKs have received corresponding APIs. Global instructions can now be loaded from the AGENTS.md file in the shared agents directory, which is set by the DSH_AGENTS_HOME variable. The official plugin catalog added the ability to install Claude Code and Codex packages on the fly, corresponding to the current release; these packages can also be used in profile configurations without an interface, SDK, or ACP. SSH has received standalone auxiliary runtimes to handle remote operations with files, processes, terminals, sandboxes, and Node PTC.

The web version can now bind to a specific IPv4 or IPv6 address and serve HTTPS directly using the new --tls-cert and --tls-key options. Other improvements include font and size settings, microphone selection for voice input, and the ability to choose when to collapse job details.

Among the fixes noted is that requests no longer hang after an idle timeout when the model stream stops, and the transport ignores cancellation, while background jobs are no longer marked as interrupted but remain active. A problem causing a failure during plugin metadata reading to block the entire request to DeepSeek has also been resolved.

The release also includes breaking changes. The mixed tool presentation mode has been removed, leaving native and ptc. Agent tools now immediately return a child identifier, requiring user SDK runtimes to support session waiting. Agent command messages are now sent directly to the target mailbox, excluding a separate outgoing mailbox and automatic retries. The default SDK profile has changed from a fixed coding agent persona to a general AI agent persona, and the bundled pi-ai library has been updated to version 1.0.2.

Similar stories

Tencent's WorkBuddy allows creation of WeChat mini-program previews from text prompts
Read more
pandaily.com

Tencent's WorkBuddy allows creation of WeChat mini-program previews from text prompts

Tencent's WorkBuddy tool can now transform a simple text prompt into a preview of a WeChat mini-program, including the publishing stage, all within a single agent, according to a report published on QQ News on October 8th.

This report is a product test conducted by a technical creator, not an official Tencent press release, and focuses exclusively on the mini-program workflow.

After updating WorkBuddy, the author activated the mini-program mode in the code tools and described a small voice memo app: speech recognition as input and text output with note sorting. The agent began developing the project based on this description. Upon completion of the build, it generated a card. Opening this card revealed a sidebar with a mini-program preview and cloud tools, including a database.

When the prompt did not contain information on what to save, WorkBuddy paused and requested clarification. During the connection process, cloud components typically outside the code generator, such as authorization, data storage, and files, were integrated. For speech-to-text conversion, the author selected Tencent's wand-asr-v1 model, and for note sorting, WorkBuddy used GLM-5.3-Flash. The report notes that the agent has a model catalog and can connect the necessary model, which occurred during this test.

Publishing has several options. A developer with a mini-program account can scan the code for binding and submit either a beta or a release build. A user who only wants to test an idea can create a trial mini-program. The author reported that the draft version of the memo app became functional in just a few minutes, with notes appearing in the database, although details were still considered incomplete.

It must be emphasized that none of these processes replace the verification or registration performed by WeChat itself. These steps remain outside the agent's scope. Binding an account or requesting a release is not equivalent to approval by WeChat. This test demonstrates a shorter path from a phrase suggestion to obtaining a preview, cloud storage, and a publication request, but it does not prove that every mini-program goes live independently.

Popular