DeepSeek Harness major update: multi-modal support is fully online, unexpected spoilers of V4 visual version trends. Only 7 days after its release, the open source agent tool DeepSeek Harness has received a new version upgrade, officially introducing enhanced multi-modal support

DeepSeek Harness major update: multi-modal support is fully online, unexpected spoilers of V4 visual version trends. Only 7 days after its release, the open source agent tool DeepSeek Harness has received a new version upgrade, officially introducing enhanced multi-modal support. With this update, the DeepSeek model adapter now supports configuration and enable native image requests, triggering widespread speculation that DeepSeek’s own visual model will be released soon. Along with the improvement of the underlying architecture, this update brings a number of practical improvements to the interactive experience. Commands such as /goal and /plan, which were previously limited to plain text transmission, now fully support sending with images; at the same time, the reference scope of @ has also been extended from simple files to historical sessions. This means that developers can simultaneously mount screenshots, local files and past conversation content when asking a single question. In order to ensure the stability of complex workflows, the new version not only implements MCP and ACP support for persistent image attachments, so that attachments will not disappear with the end of a single call, but also successfully fixes the request failure problem caused by excessive image size and excessive cumulative load of historical images in multiple rounds of conversations. In addition to the upgrade of multi-modal capabilities, the new version also takes frequent actions in component distribution and architecture optimization. The Claude Code and Codex are no longer distributed with the principal and are instead converted to Profile Bundles that are installed on demand. Among them, Codex has ushered in two major new features: first, it supports non-interactive permission mode, allowing it to run smoothly in unattended processes; second, it supports multiple named instances, allowing multiple Codex sub-agents with different configurations and names to exist at the same time. At the same time, the "return path" of the child agent has also been optimized. reportDelivery can now provide timely feedback and wake up the parent task, eliminating the trouble of the parent task blindly waiting for the child task to end. In terms of other functional optimizations, web_search has supported concurrent queries, Windows PTY terminal has officially added support for persistent PowerShell sessions and is enabled by default in the minimalist mode preset. The read, write and fork performance of the SQLite backend has also been significantly improved, and the storage volume has been further reduced. In addition, the official confirmed that "DeepSeek Harness" has been registered as a trademark, and the relevant brand usage regulations have also been updated simultaneously. via AI News (author: AI Base)