DeepSeek harness receives major update with enhanced multimodal support

DeepSeek Harness, a framework developed by DeepSeek AI, has released a significant update that expands its multimodal capabilities. The new version introduces native support for image requests through the DeepSeek model adapter, allowing users to configure and enable visual interactions. This update also extends the functionality of existing commands such as /goal and /plan, enabling them to accept image inputs alongside text. Additionally, the MCP and ACP modules now support persistent image attachments, ensuring that images remain available across multiple interactions rather than being discarded after a single call. The update also resolves issues related to image size limitations and excessive payload accumulation in multi-turn conversations, which previously caused model request failures. Furthermore, the update separates Claude Code and Codex from the main distribution, making them optional installable profiles. Codex now includes non-interactive mode for automated workflows and multiple named instances for managing different configurations. The reporting mechanism for subagents has been improved, allowing for real-time feedback and task resumption without requiring the parent task to wait. Windows PTY terminals now feature persistent PowerShell sessions, with the minimal mode preset to enable these sessions by default. The minimal mode retains only persistent bash and file editor tools. SQLite backend performance has been optimized, with reduced storage size, though the data structure is no longer compatible with previous versions. Finally, 'DeepSeek Harness' has been officially registered as a trademark, with detailed usage guidelines provided in the brand usage document.