dsh-vision-bridge

Paste an image into the DeepSeek Harness chat box and let the agent see and describe it automatically — no more "current model does not support images" errors, no manually handing over file paths.

国内镜像 / Mirror: also hosted on gitee.com/chill109/dsh-vision-bridge — Gitee is a mainland-China Git host (faster access from mainland China; use it if GitHub is slow).

中文版见 README.zh.md。

Part 1 — If you are a human

This part is written for people who just want the feature installed.

What it does

Before this project, pasting an image into the chat box with a text-only model (e.g. deepseek-v4-flash) showed:

当前模型不支持图片，请切换支持图片的模型 / The current model does not support images; switch to a model that does

After installing this project:

1. Paste any image into the send box (it shows as a thumbnail, as usual).
2. Send it — the message goes through even though the chat model cannot take images.
3. The harness automatically hands the agent the image as a local file path, and the agent automatically calls whatever vision tool it has available (e.g. vision_chat from qwen-mm-plugins-api) to recognize it — it answers with a description, OCR, VQA… whatever you asked.

You never type a path; the recognition happens on its own.

Please understand this clearly: this project is only a bridge — it makes the image sendable and hands it to the agent as a path. It does not include or install any vision model:

• The vision plugin must be configured by you (e.g. register qwen-mm-plugins-api in profiles\web\cordis.patch.yml with a working API key);
• The vision model is freely selectable — any tool that takes an image and returns text works;
• If you did not specify which vision model to use, let the AI recommend a compatible one (e.g. mcp-qwen-mm-plugins-api / vision_chat); the placeholder tells the agent to use whichever vision tool it currently has.

What this project is NOT: it has no vision capability of its own

dsh-vision-bridge ships no vision model, no vision tool, and no image-understanding code. It only does two things:

1. Allows text-only main models (e.g. deepseek-v4-flash) to send images in a session (removes the harness rejection checks);
2. At request time, projects image blocks into "attachment path + vision instruction" text so the main model calls whatever vision tool you have registered to read the image.

The division of labor: the main model (DeepSeek, etc.) thinks, the vision tool sees — and the "seeing" side is not in this repository. Without a vision tool, this bridge itself cannot see anything.

Recommended pairings (the vision side is yours to set up)

Vision tools (MCP plugins) — qwen-mm-plugins is the recommended pairing:

• api capability: cloud vision tools vision_chat (describe/VQA), ocr (text extraction), grounding (bounding boxes), and more; register mcp-qwen-mm-plugins-api in profiles\web\cordis.patch.yml and the bridge's placeholder instruction will call it automatically;
• core capability: local reading/visualization (images, video, documents, code, etc.) — no network needed.

Free vision APIs (pick any; any OpenAI-compatible endpoint works):

Example with Zhipu — in the qwen-mm-plugins config file ~/.qwen-mm-plugins/config:

DASHSCOPE_API_KEY = <Zhipu API key (get it for free at open.bigmodel.cn)>
DASHSCOPE_BASE_URL = https://open.bigmodel.cn/api/paas/v4
QWEN_MM_VISION_MODEL = glm-4v-flash

Tip: glm-4v-flash caps output at 1024 tokens (qwen-mm-plugins adapts automatically); the free tier is rate-limited to roughly 3-10 RPM — plenty for chat use.

Requirements

• DeepSeek Harness web UI (dsh web --host 127.0.0.1 --port 3080)
• The qwen-mm-plugins-api MCP server registered in your web profile (provides the vision_chat tool; needs a working OpenAI-compatible API key — Zhipu / SiliconFlow / DashScope all work)
• git (used by the installer to apply the patches)

Manual deployment (Windows)

1. Apply the host patch — removes the "model does not support images" rejection:

   cd "$env:USERPROFILE\.dsh\profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib"
   git -c core.autocrlf=false apply --unsafe-paths --directory="$PWD" "..\..\..\..\..\..\dsh-vision-bridge\patch\dsh-host-apiproxy.patch"

2. Apply the LLM patch — text-only models receive images as a path + a vision instruction:

   cd "$env:USERPROFILE\.dsh\profiles\node_modules\@deepseek-ai\dsh-llm-pi-ai\lib"
   git -c core.autocrlf=false apply --unsafe-paths --directory="$PWD" "..\..\..\..\..\..\dsh-vision-bridge\patch\dsh-llm-pi-ai.patch"

3. Restart the harness — close the DeepSeek Harness Server window, run start-dsh.bat (or restart dsh web), then hard-refresh the browser (Ctrl+F5).

One-command alternative

.\install.ps1          # applies all four patches and restarts the harness

Let an AI deploy it

Copy/paste this whole paragraph into your AI assistant:

Deploy the dsh-vision-bridge feature from https://github.com/chou109/dsh-vision-bridge (mainland-China mirror: https://gitee.com/chill109/dsh-vision-bridge) to my local DeepSeek Harness web profile.

1. Windows: run .\install.ps1 from the repo root — it applies patch\dsh-host-apiproxy.patch, patch\dsh-host-apiproxy-selectmodel.patch, patch\dsh-llm-pi-ai.patch and patch\dsh-llm-deepseek.patch into %USERPROFILE%\.dsh\profiles\node_modules\@deepseek-ai\... with git -c core.autocrlf=false apply, then restarts the harness. macOS/Linux: ./install.sh (restart dsh web yourself).
2. If the scripts are unavailable, apply the four patches manually exactly as the README "For AI" section describes.
3. Note: this project is only a bridge between "image sending" and a vision-model plugin — it does not include a vision model. The vision plugin must be configured separately and the vision model can be freely chosen. If I did not specify which vision model to use, please recommend a compatible one (e.g. mcp-qwen-mm-plugins-api / vision_chat), and confirm it is registered in profiles\web\cordis.patch.yml with a working API key.
4. After the harness is back, verify: I paste an image into the chat box and send it — the agent must automatically describe the image (it receives [图片附件：<path>（如需了解内容，请调用你可用的视觉识图工具…）] and automatically calls its available vision tool).

Usage

1. Paste (Ctrl+V) or drag an image into the chat box.
2. (Optional) type a question, e.g. "这是什么".
3. Send. The agent automatically recognizes the image with vision_chat and answers.
4. Pasted images also work with no text at all — the message auto-sends.

Uninstall

.\install.ps1 -Uninstall

or manually: git apply -R all four patches (or reinstall @deepseek-ai/dsh-host-apiproxy@0.1.0-rc.6, @deepseek-ai/dsh-llm-pi-ai@0.1.0-rc.6 and @deepseek-ai/dsh-llm-deepseek), then restart the harness.

Part 2 — If you are an AI

This part is written for AI agents that install, debug, or extend this feature. It assumes you can run shell commands and read the dsh packages in node_modules.

What this is (facts)

Paste-to-vision is four patches across three shipped dsh packages (no client-side change needed — the chat composer already allows pasting; the block was a server-side rejection at send time):

1. patch/dsh-host-apiproxy.patch — in @deepseek-ai/dsh-host-apiproxy/lib/index.js, the prompt RPC handler previously rejected any message containing image parts when the selected model's inputModalities lacked image (returning attachment-error / MODEL_DOES_NOT_SUPPORT_IMAGES, which the UI renders as "当前模型不支持图片…"). The patch deletes that rejection: image parts are admitted for any model. (1 hunk; the file gets smaller.)

2. patch/dsh-llm-pi-ai.patch — in @deepseek-ai/dsh-llm-pi-ai/lib/index.js, the stream() entry previously threw UNSUPPORTED_CONTENT when a text-only model received image blocks. The patch instead projects each image block to a text placeholder:

   [图片附件：<abs path>（如需了解内容，请调用你可用的视觉识图工具识别此图片；例如 vision_chat，images 参数传此路径）]

   The path is resolved by imageAttachmentPath() from the content-addressed ref: <DSH_HOME>/attachments/v1/objects/<aa>/<sha256>.<ext> (a best-effort hardlink adds the extension so tools can sniff it). The projection recurses into tool-result content too, so images returned by tools (e.g. read_image) are handled the same way. (3 hunks.)

3. patch/dsh-host-apiproxy-selectmodel.patch — the same package's selectModel RPC previously rejected switching to a text-only model when the session already contained images (model-unavailable: "does not accept image input, but this session already contains images"). The patch removes that rejection: image sessions may switch to any model — the images are bridged at request time by the projection patches below. (1 hunk.)

4. patch/dsh-llm-deepseek.patch — newer harness versions route the DeepSeek official provider (deepseek-official, e.g. deepseek-v4-flash) through the dedicated adapter @deepseek-ai/dsh-llm-deepseek, which unconditionally threw UNSUPPORTED_CONTENT ("The DeepSeek chat-completions adapter does not support image content.") for image blocks. The patch applies the same projection as pi-ai at the request() entry, recursing into tool-result content. (3 hunks.)

5. The recognition side is NOT part of this repo — the placeholder is a generic instruction ("use whatever vision tool you have available"), so the agent automatically picks the vision tool currently registered in the profile. A typical pairing is the vision_chat tool of the qwen-mm-plugins-api MCP server (registered in profiles/web/cordis.patch.yml → mcp-qwen-mm-plugins-api with a working API key), but any tool that accepts an image path and returns text works.

Data flow: paste → draft thumbnail → send → prompt RPC admitted (patch 1) → message persisted with image parts (UI history still shows the image) → LLM request serialization hits a text-only model → projection (patch 2 / patch 4, whichever adapter the route uses) → agent context receives the path placeholder → agent calls vision_chat(path) → answer. Switching models is allowed on image sessions too — selectModel no longer rejects them (patch 3).

Deployment (exact steps)

$profiles = "$env:USERPROFILE\.dsh\profiles"          # or $env:DSH_HOME\profiles

# 1. host-apiproxy (removes the rejection)
$d = "$profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" patch\dsh-host-apiproxy.patch

# 2. llm-pi-ai (image -> path text projection)
$d = "$profiles\node_modules\@deepseek-ai\dsh-llm-pi-ai\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" patch\dsh-llm-pi-ai.patch

# 3. host-apiproxy (model-switch rejection removal)
$d = "$profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" patch\dsh-host-apiproxy-selectmodel.patch

# 4. llm-deepseek (DeepSeek adapter: image -> path text projection)
$d = "$profiles\node_modules\@deepseek-ai\dsh-llm-deepseek\lib"
git -c core.autocrlf=false apply --unsafe-paths --directory="$d" patch\dsh-llm-deepseek.patch

• Patches target index.js with a/index.js/b/index.js headers; run from the repo root.
• Line endings: bundles are LF-only; -c core.autocrlf=false is mandatory on Windows.
• Version pin: context is exact for 0.1.0-rc.6. If git apply fails, the installed version differs — re-diff against npm pack @deepseek-ai/dsh-host-apiproxy@0.1.0-rc.6 (and the same for dsh-llm-pi-ai and dsh-llm-deepseek).
• Idempotency: install.ps1/install.sh detect the patch state via markers (host: string MODEL_DOES_NOT_SUPPORT_IMAGES absent ⇒ patched; selectmodel: does not accept image input absent ⇒ patched; llm (pi-ai / deepseek): function projectImageBlocksToText present ⇒ patched).

Verification (after deploy + restart + hard refresh)

1. Paste an image into the chat box and send it (empty text auto-sends).

2. Expected: the agent's turn shows it analyzed the image (description/OCR/answers).

3. On the wire: the user message that reached the LLM contains [图片附件：<path>（如需了解内容，请调用你可用的视觉识图工具识别此图片；例如 vision_chat，images 参数传此路径）] — the placeholder is a generic instruction (it does not hardcode one tool); the agent calls whichever vision tool it has. This is by design, not an error.

4. Direct checks:

   # host patch live: the reason string is gone from the running code
   Select-String "$profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib\index.js" -Pattern 'MODEL_DOES_NOT_SUPPORT_IMAGES'   # -> no match
   # llm patch live (pi-ai adapter)
   Select-String "$profiles\node_modules\@deepseek-ai\dsh-llm-pi-ai\lib\index.js" -Pattern 'projectImageBlocksToText'            # -> match
   # llm patch live (deepseek adapter)
   Select-String "$profiles\node_modules\@deepseek-ai\dsh-llm-deepseek\lib\index.js" -Pattern 'projectImageBlocksToText'         # -> match
   # selectmodel patch live: the reason string is gone from the running code
   Select-String "$profiles\node_modules\@deepseek-ai\dsh-host-apiproxy\lib\index.js" -Pattern 'does not accept image input'   # -> no match

Common failure modes

Operations

• Restart harness: taskkill /F /T /PID <node dsh web pid> then npx -y @deepseek-ai/dsh web --host 127.0.0.1 --port 3080. install.ps1 does this automatically.
• Rollback: install.ps1 -Uninstall (or git apply -R both patches), restart.
• Junction caveat: profiles\node_modules\@deepseek-ai\* are junctions to the npx cache. Re-running npx -y @deepseek-ai/dsh with a newer package version can overwrite patched files — re-apply after upgrades.
• No client changes: do not patch dsh-client-ui-conversation for this feature; the composer already allows pasting.

Extra — how it works (and why it's shaped this way)

• The rejection lived server-side, not client-side. The composer accepts pasted images for any model; the error appeared because the host's prompt RPC checked modelInfo.inputModalities and refused. Removing that one check is the entire client-visible fix.
• Why path text and not pixels? A text-only LLM cannot consume image bytes. The image is stored content-addressed by the harness anyway (attachments/v1/objects/…), so the cheapest reliable bridge is a path. The placeholder text doubles as an instruction, so any agent with the vision_chat tool recognizes the image without extra configuration.
• Images from tools work too. The projection recurses into tool-result content, so read_image / save_view results are bridged the same way for text-only models.
• Attachment display is preserved. Only the LLM serialization is rewritten; the persisted message and the chat UI keep the real image block.

FAQ

• Q: Is my image sent anywhere extra? The pasted image is stored locally by dsh and read by the vision tool via its local path; the vision model (e.g. mimo-v2.5 via your DashScope-compatible endpoint) receives it exactly as before — the feature only changes which component reads the image (the vision model instead of the chat model).
• Q: Does it work with image-capable models? Yes, unchanged: when model.input includes image, no projection runs and the model sees the pixels directly.
• Q: Why is the placeholder in Chinese? The agent prompt convention in this profile is Chinese; the text is opaque to the model anyway — it only needs to identify the path and the tool name.
• Q: The repo has no LICENSE. Add one before publishing (MIT suggested).

Made for DeepSeek Harness users who want to paste a picture and get an answer.