DeepSeek published DeepSeek Harness v0.2.1-rc1 on October 17, the first release candidate of the 0.2.1 series for its open-source agent runtime. The release notes on GitHub bundle the main user- and developer-facing changes made since v0.2.0-rc.2. Several of them, including reasoning-trace translation, the Git Worktrees plugin and the Claude Code Mods compatibility layer, already shipped in the alpha.1 and alpha.2 pre-releases that Pandaily covered earlier. This article focuses on what is new relative to alpha.2.
The biggest addition is around plugins. Users can now discover and install plugins from custom plugin index sources rather than only the official list. The plugin manager gains a 'let the Agent create a plugin' entry that keeps a draft in a creation mode and only starts work once the user sends a request. Enabling, disabling, uninstalling and updating plugins now syncs dependencies, and when hot reload is on, an updated or reinstalled replaceable plugin takes effect immediately; otherwise the app prompts for a restart. Optional plugins support a minimal mode, and the official list can install Claude Code and Codex bundles matched to the current release for headless, SDK and ACP profiles. An experimental Session state interface lets plugins write and read their own session records.
Model and context handling also moved forward. Custom model APIs can have reasoning levels set per model in the interface instead of a config file. The pi-ai router handles in-session system prompt updates and dynamic tool addition and removal according to model capabilities. Minimal mode can read images and automatically compress long conversations, and the web and desktop apps add a /compact command for manual compression. Global instructions can read AGENTS.md from a shared agents directory set through DSH_AGENTS_HOME.
For developers, a new working_directory tool arrives alongside TypeScript and Python SDK interfaces for reading and switching a session's working directory. The web version can sign in to a DeepSeek account over local HTTP, secrets and tokens detected in sent messages are masked by default, new sessions can be prefilled with an unsent prompt, and Markdown previews render YAML front matter as a field list.
Separately, an automated tracking account on X reported that an entry labeled 'deepseek-v4.1-flash-1017' briefly appeared in DeepSeek's official API model list before being removed. The suffix follows the dated naming of earlier snapshots such as V4-Flash-0731 and V4-Pro-0813. The same account noted that the public deepseek-flash model still points to V4.1 Flash, not a 1017 build. DeepSeek has not commented on whether such a snapshot exists or what it would change.
SEE ALSO: DeepSeek Harness v0.2.1-alpha.2 Adds Reasoning Translation and Git Worktrees · DeepSeek Harness v0.2.1-alpha.1 Adds Claude Code Mods Compatibility Layer · DeepSeek Adds Standalone Search and Fetch APIs