DeepSeek Harness advanced through several release candidates and alpha versions in August, adding multimodal capabilities, improved subagent management, and enhanced conversation navigation.
Models and Capabilities
- DeepSeek-V4-Flash-Vision-Exp model: A new multimodal model is now available in the DeepSeek adapter, enabling visual understanding for image analysis tasks.
- Configurable image input: Native image requests for DeepSeek adapters and image input for commands like /goal and /plan are now configurable, with images rendering immediately during upload.
Agents and Automation
- Subagent management: Claude Code and Codex subagents are now installable on demand as Profile Bundles with non-interactive permission modes and named instances, and plugin-controlled provider sign-in lets agents select provider, model, and reasoning effort with authorization checks.
- Agent Preset switching: Users can now switch between Agent Presets and perform cross-preset search within conversations.
Interface and Workflow
- Question history display: Question history now displays as readable Q&A cards showing cancelled and interrupted states for better conversation tracking.
- Conversation navigation: Users can preview and jump to any turn in paginated conversations directly from right-hand navigation, even if those turns have not been loaded yet.
- Context references in @ menu: File and session references are now available in the @ menu for easier context selection.
Platform and Integration
- Full ACP support: Comprehensive support for session controls, model settings, MCP, permissions, and cancellation is now available.
- Plugin organization and WebFetch: Plugins are now grouped by conversation and global scope, and public WebFetch is enabled by default with SSRF protection.