DeepSeek Harness release candidate adds V4.1-Flash-1017 and subagent teams

DeepSeek released dsh-v0.2.2-rc.2, the first release candidate in the 0.2.2 series of its open-source coding-agent harness, on November 11, 2026. The update was published on the official deepseek-harness GitHub repository at 11:33 UTC and follows the prior v0.2.1 series.

What changed in this release candidate

• Released as tag dsh-v0.2.2-rc.2 on the official deepseek-ai/deepseek-harness GitHub repository, published November 11, 2026 at 11:33 UTC.
• Adds support for the DeepSeek-V4.1-Flash-1017 model.
• Adds new DeepSeek Search and Fetch APIs, togglable in Web Search settings, with web search now defaulting to the DeepSeek Search API.
• Git Worktree now supports reviewing changes by branch, commit or file category together with GitHub reviews.
• SubAgent gains named, multi-level collaboration with shared team panels per session, and AgentTeam adopts unified named subagents.
• MCP services can now be dynamically registered or removed by plugins, and session archiving automatically cleans up Git worktrees by default.

What it means for people building with AI

DeepSeek Harness is the official open-source agent and coding CLI tied to DeepSeek's models, so these additions change how developers structure agentic workflows directly on top of them. Named, multi-level subagent collaboration with shared team panels lets teams coordinate several subagents inside one session instead of running isolated single-agent tasks. Letting plugins dynamically register or remove MCP services means tool integrations can be added or dropped without restarting a session, which matters for anyone building multi-agent coding workflows on DeepSeek models.

Availability

This is a release candidate, available now as tag dsh-v0.2.2-rc.2 on the deepseek-ai/deepseek-harness GitHub repository. The desktop app also lets users choose between stable and development update channels.

Sources

• deepseek-harness dsh-v0.2.2-rc.2 (DeepSeek (GitHub), November 11, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.