Before installing, read the pinned README requirements and compare them to your DSH profile, Node runtime, Python runtime, package-cache policy, browser availability, and endpoint policy. The source documents both Web and Headless profile commands, but they are not pinned to a release; this page therefore treats the source as instructions to review, not as a command to copy blindly.
BundleVision / Developer toolsv0.1.2Developer Preview
DSH Vision Toolkit
DSH Vision Toolkit is a native DSH bundle for structured visual analysis, OCR, grounding, screenshots, artifact output, and Web settings.
279 repository starsMIT Updated Aug 14, 2026
Review snapshot
- Manifest
- Passed
- Install
- Not tested
- Runtime
- Not tested
Static compatibility snapshot dated 2026-08-14; install and runtime remain independent checks.
Install / upstream documentation
Use the pinned upstream documentation
No version-pinned install command is recorded for this listing. Use the pinned README as the source of truth; AgentPlugin.net does not invent missing steps, Profiles, or recovery instructions.
The pinned README documents unpinned package commands. It also contains a repository/package version mismatch at this commit (package.json is 0.1.2 while the README calls 0.1.4 current), so this listing deliberately links to the pinned instructions rather than supplying a command as reproducible release guidance.
Source: Pinned upstream install and lifecycle sectionPlugin brief
What DSH Vision Toolkit adds to DeepSeek Harness
DSH Vision Toolkit is a DeepSeek Harness plugin distributed as @dsh-external/dsh-vision-toolkit. The pinned package.json declares a DSH bundle patch, and the pinned YAML patch inserts the vision-toolkit bundle row. Upstream describes a native integration of agent-vision-toolkit with DSH lifecycle management, structured tool schemas, Credentials, Artifacts, Settings, and Web presentation. It is aimed at visual engineering work rather than at making every chat model natively multimodal.
The README lists visual operations for image questions, OCR, original-pixel grounding, detection, cropping, SVG tracing, pixel differences, foreground extraction, colors, and local HTML screenshots. These capabilities can support screenshot-based debugging or UI restoration, but their output is evidence to inspect, not a substitute for visual review. Some operations are local while others require an OpenAI-compatible vision endpoint and a DSH Credential, which changes the trust and data-flow boundary.
The package documents a managed runtime and an external runtime. Managed mode validates a packaged upstream snapshot, prepares an isolated Python environment, and uses a requirements lock; external mode accepts only an exact upstream pin and rejects modified or untracked checkout content according to the README. Those are useful design claims to inspect, but they are not an AgentPlugin.net runtime certification. The first activation may need network access or an available package cache.
The artifact model is especially relevant to DSH users. Upstream says output is written below the session workspace, with validation before commit and optional signed preview/download capability URLs when a Web HTTP host exists. Keep workspace access, artifact retention, browser preview availability, and any remote vision credential under your own policies. Sensitive images, OCR text, and generated artifacts may all be valuable project data.
Evaluation workflow
Evaluate DSH Vision Toolkit with clear boundaries
Start with local, non-sensitive images and a disposable workspace. Confirm which tools are expected to work without a credential and which will contact the configured OpenAI-compatible endpoint. Then inspect the session log and artifact directory. The README’s explicit distinction between local health checks and connection testing is useful: a connection test sends the configured Credential to the endpoint, while an ordinary settings read should not.
Configure the DSH Credential reference rather than embedding a secret in a profile patch or prompt. Upstream says the setting stores the reference and resolves it only for remote subprocess work. Validate that assertion in your deployment and apply your own secret-management, egress, logging, and incident-response controls; a repository statement is not a replacement for organization-level controls.
For UI restoration and screenshot comparison, use an independent human visual check after the tool result. The README says HTML screenshot rendering disables network access in its pinned adapter and uses a disposable Chrome-family profile, but browser discovery and local-file scope still deserve review. Keep source images and generated HTML within allowed workspace roots, and treat any text inside images as untrusted content rather than as instructions.
Verification ledger
Evidence, not installation claims
Manifest
PassedPinned package.json declares dsh.bundle.patch as ./cordis.patch.yml and the pinned patch inserts id vision-toolkit with name @dsh-external/dsh-vision-toolkit.
Checked Aug 14, 2026
Install
Not testedNo clean Web or Headless DSH install was run.
Runtime
Not testedNo managed-runtime, Credential, Artifact, or visual tool test was run.
| Layer | Result | What was checked |
|---|---|---|
| Manifest | Passed | Pinned package.json declares dsh.bundle.patch as ./cordis.patch.yml and the pinned patch inserts id vision-toolkit with name @dsh-external/dsh-vision-toolkit.[Aug 14, 2026] |
| Install | Not tested | No clean Web or Headless DSH install was run. |
| Runtime | Not tested | No managed-runtime, Credential, Artifact, or visual tool test was run. |
Source or manifest review can be recorded while Install and Runtime remain not-tested. Treat these as separate verification layers.
Execution surface
Prepare, build, and code-execution review
prepare
not recorded
No review record is available for this execution surface. Inspect the upstream repository before running it.
build
notice recorded
The package prepack script runs a build; source or package workflows can execute build tooling. Review the pinned repository before using non-registry installation paths.
code execution
review recorded
Managed runtime preparation uses Python tooling and dependency installation; inspect locked requirements, package sources, and workspace permissions.
network
review recorded
Remote visual tools can send image-derived requests to the configured OpenAI-compatible endpoint and may need package downloads during first managed preparation.
configuration
review recorded
Credentials, allowedDirs, runtime mode, external path pinning, artifact storage, and optional Web previews define the plugin’s effective boundary.
Third-party code can execute during dependency preparation, build, or runtime. Read the upstream source and its changes before use.
Declared scope
Capabilities and prerequisites
Capabilities
- Provides ten declared visual-tool schemas including OCR, grounding, crop, trace, pixel diff, and HTML screenshots
- Uses DSH Credentials and Settings for an OpenAI-compatible vision endpoint
- Writes validated artifacts under the workspace and supports Web previews when an HTTP host is available
- Supports managed and strictly pinned external upstream runtime modes
Prerequisites
- DeepSeek Harness Web or Headless profile with pnpm
- Node.js ^22.19.0 or >=24.0.0, as declared in package.json
- Python 3.11 or newer for the managed runtime
- Network/package-cache access for first preparation when dependencies are not cached
- A DSH Credential and compatible endpoint for remote vision tools; Chrome-family browser only for HTML screenshots
Frequently asked questions
DSH Vision Toolkit FAQ
What is DSH Vision Toolkit?
It is a native DeepSeek Harness bundle for structured visual tasks such as OCR, grounding, image analysis, screenshot comparison, artifact output, and Web settings.
Is it a DSH Bundle?
Yes. The pinned package.json declares dsh.bundle.patch and the pinned cordis patch inserts id vision-toolkit with package @dsh-external/dsh-vision-toolkit.
Why does this listing omit an install command?
The upstream commands are unpinned, and the pinned README’s stated current version differs from package.json. The exact README is linked so users can reconcile the release before installation.
Does every tool need a remote vision service?
No. The README distinguishes local operations from tools that require an OpenAI-compatible vision endpoint and DSH Credential. Review each tool and configuration before use.
Where are generated artifacts written?
Upstream says artifact-producing tools write under <workspace>/.dsh-vision-toolkit/artifacts and validate outputs before commit. Check your actual workspace permissions and retention policy.
Has AgentPlugin.net tested the bundle?
No. Static package, patch, README, and license evidence were inspected; no installation, credentials flow, runtime preparation, or visual tool smoke test was performed.