AnswersPublished24 sources
DeepSeek Harness v0.1.0-rc.8: 14 Changes That Make Images First-Class Agent Input
DeepSeek Harness v0.1.0 rc.8 was published on August 19, 2026—not August 21—and its headline change is image transport: compatible vision models can receive native images, while text only models get structured OCR and... The 14 update release also adds mixed text and image input to /goal and /plan, file and session...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- DeepSeek Harness v0.1.0 rc.8 was published on August 19, 2026—not August 21—and its headline change is image transport: compatible vision models can receive native images, while text only models get structured OCR and...
- The 14 update release also adds mixed text and image input to /goal and /plan, file and session references through the @ menu, on demand Claude Code and Codex subagents, persistent Windows PowerShell sessions, and fix...
- Tool layer vision is useful for screenshots, slides, diagrams, and text heavy graphics, but it is not a substitute for native vision on photos, subtle visual context, or complex scene understanding.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What changed in DeepSeek Harness v0.1.0-rc.8, released on August 21, 2026 as its first major post-beta update after the August 13 public bet. Article summary: The release was published on August 19 UTC, although coverage around August 20–21 described it as the first major update following the August 13 public beta/open-source launch. Its main change was to make images first-cl. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
DeepSeek Harness v0.1.0-rc.8 is primarily an orchestration and input upgrade, not proof that every DeepSeek model can see. The release was published on August 19, 2026, six days after the public beta and open-source launch; some coverage dated the announcement or discussion to August 20–21. 9
16 Its 14 changes make images usable throughout agent workflows while also improving subagents, terminal behavior, tool execution, storage, and compatibility.
4
The headline change: images enter the agent loop
RC.8 adds a configurable native-image request path to the DeepSeek model adapter. When the selected model or backend supports image input, Harness can pass the image directly to it rather than treating the attachment as separate from the task. 2
8
That image support reaches core planning commands:
/goalaccepts mixed text-and-image input./planaccepts mixed text-and-image input.- The
@menu can reference local files and earlier sessions as work continues.2
12
15
This matters for coding-agent workflows because a screenshot, interface mock-up, diagram, or error state can become part of the goal and plan instead of requiring a user to describe it manually.
The important caveat is that RC.8 adds the transport and orchestration path; it does not automatically give a text-only model independent visual understanding. Native interpretation depends on the configured model or compatible backend. 8
9
What the 14 updates cover
Claude Code and Codex become installable subagents
Claude Code and Codex can be installed as on-demand profile bundles within the subagent system. Codex also gains a non-interactive permission mode and support for multiple named instances, making it possible to run separately configured Codex subagents in one workflow. 2
4
Windows terminals keep shell state
Windows PTY terminals gain persistent PowerShell sessions. The behavior is enabled by default in the minimal preset, so successive commands do not need to recreate the shell state each time. 2
Tool execution becomes more capable
The release improves tool calling, including concurrent web_search, waking the parent task after a subagent delivers a report, and automatically opening a local browser for dsh web. It also improves performance for large-session forks and SQLite storage. 2
4
Image, streaming, and gateway failures are addressed
RC.8 fixes request failures associated with an oversized single image or too many images accumulated in conversation history. It also preserves reply prefixes when a streaming response is cancelled and the conversation is continued or forked. Custom OpenAI-compatible gateway interoperability is improved for differences in request formats and missing inference-content returns. 2
4
Together, these changes make the release broader than a simple image-upload feature: images can participate in planning, session history, and delegated work, while the runtime becomes more resilient around long or tool-heavy sessions. 2
13
How “tool-layer vision” works for text-only models
When a selected model does not declare image-input support, Harness can use tools to turn an image into structured textual evidence. The reported pipeline includes OCR text with coordinates, color-ratio statistics, selected pixel-row scans, image dimensions, and color-mode metadata. That evidence is then supplied to the text model for reasoning. 1
2
The model is not receiving the original visual field in the same way a native vision model does. It is reasoning from measurements and extracted descriptions—a useful fallback, but a lossy one.
Where the fallback is strongest
Tool-layer vision is best suited to images with recoverable structure, including:
- Application and user-interface screenshots
- Slides and presentation screenshots
- Flowcharts, diagrams, and simple graphics
- Text-heavy images with clear layouts and strong contrast
These tasks often depend on readable text, alignment, boundaries, colors, or other signals that OCR and targeted pixel inspection can capture.
Where native vision is still necessary
The approach is substantially weaker for natural photographs, fine-grained object recognition, visual style, subtle context, and complex spatial or scene relationships. OCR and sampled pixel statistics cannot reliably reconstruct every detail needed for those tasks. 1
2
The practical rule is straightforward: use a native vision model for genuine image understanding, and use tool-layer vision for targeted evidence extraction or as a fallback when the main model is text-only. It can help extract text, inspect dominant colors, locate simple boundaries, or validate pixel-level interface differences, but its output should not be treated as equivalent to native visual reasoning. 2
3
Community plugins that extend image workflows
Community tooling had already explored several ways to give text-only Harness agents access to visual evidence. Reported options include dsh-vision, dsh-vision-toolkit, and modlens; plugin directories also list local OCR and pseudo-vision bridges. 10
23
Two examples illustrate the design space:
- ModLens returns structured evidence covering OCR, layout, entities, and relationships for text-only models.
18
dsh-vision-toolkittargets long-screenshot OCR, visual grounding, UI reconstruction, and pixel comparison.21
These plugins complement rather than replace RC.8’s native-image path. A compatible vision model can receive the image directly; a plugin can provide specialized OCR, layout, pixel, or external-model tooling; and a text-only model can reason over the resulting evidence. That separation lets developers choose between richer native perception and more targeted, inspectable tool output.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "DeepSeek Harness v0.1.0-rc.8: 14 Changes That Make Images First-Class Agent Input"?
DeepSeek Harness v0.1.0 rc.8 was published on August 19, 2026—not August 21—and its headline change is image transport: compatible vision models can receive native images, while text only models get structured OCR and...
What are the key points to validate first?
DeepSeek Harness v0.1.0 rc.8 was published on August 19, 2026—not August 21—and its headline change is image transport: compatible vision models can receive native images, while text only models get structured OCR and... The 14 update release also adds mixed text and image input to /goal and /plan, file and session references through the @ menu, on demand Claude Code and Codex subagents, persistent Windows PowerShell sessions, and fix...
What should I do next in practice?
Tool layer vision is useful for screenshots, slides, diagrams, and text heavy graphics, but it is not a substitute for native vision on photos, subtle visual context, or complex scene understanding.
Sources
agenccy.aiDeepSeek Shipped Image Input for Its Agent CLI. Its Own Models Still Cannot See.
dsharness.orgThe DeepSeek Harness changelog
news.qq.comDeepSeek Harness首发新版本：14项更新，多模态能力拉满_腾讯新闻
cctest.aiDeepSeek Harness v0.1.0-rc.8 Adds Multimodal Agent Input - CCTest