DeepSeek Harness adds image input in first post-beta update

A screenshot lands in a developer's task, and the model has no vision capability. In DeepSeek Harness v0.1.0-rc.8, the job no longer has to stop there: the first major update since the public beta adds image input, subagent upgrades and 14 changes across the agent workflow.

The release arrived less than a week after DeepSeek Harness opened its public beta and open source on August 13. On beta night, developer Zhidx pulled the code, translated an 88-page paper and built a Snake game. Now models with vision can receive images directly, while the /goal and /plan commands accept text and images together.

The mechanism is more interesting when the underlying model cannot see. Developer Yinsen found that a failed direct read_image call can trigger a fallback using optical character recognition, color-ratio statistics, pixel-row scanning, image dimensions and color-mode metadata. Harness passes those structured clues to a text model, which can reconstruct useful information from presentation slides, flowcharts and interface screenshots. Real photos are less reliably recovered.

The same update widens the team around the main model. Claude Code and Codex can be installed as Profile Bundles and used as subagents; Codex adds a non-interactive permission mode and multiple named instances. Windows users get persistent PowerShell sessions in the PTY terminal, while fixes address failed image requests, interrupted streaming and custom OpenAI-compatible gateways that used different request formats.

So what changes in practice? A developer can bring a screenshot, local file or previous session into one task, ask an image-capable model to inspect it, or let a text-only model work from extracted visual signals. That makes the harness useful across more models without requiring every base model to support vision. The limits are visible: the software is still a public beta release candidate, and the fallback works best on clear, structured images rather than photographs.