MLView

Ask the assistant in your editor to explain an ML workflow, then explore its answer as an interactive diagram linked to the source. MLView is a skill for GitHub Copilot, Codex and Claude Code, a document format, and a VS Code viewer. The assistant reads the code and configuration and writes the diagram: steps, connections and findings, with quotes from the files it read.

Preview. The format and the viewer can still change between releases, and fixes go into the next release.

The diagram of samples/configured_training, with a step selected.

How it works

flowchart LR
    A[Your assistant + MLView skill] --> B[Read source and config]
    B --> C[LLM interpretation]
    C --> D[WorkflowDocument draft]
    D --> E[Local validation and publication]
    E --> F[Interactive VS Code diagram]
    F --> G[Source navigation]
    A -->|Explicit refinement| C

Loading

• MLView needs no model API key of its own. Your assistant supplies the model, its tools and its permissions.
• The skill's helper checks a draft's structure, cited paths and exact quotes, records file hashes, and publishes a *.mlview.json file.
• Neither the helper nor the viewer imports or runs your code.
• Opening, filtering and navigating a diagram never call a model.

Install

Each release attaches the extension (a .vsix file) and the two skill ZIPs to its page under Releases.

• Download the .vsix file. Install it with VS Code's Extensions: Install from VSIX... command.
• Download the skill ZIP for your assistant and extract it at the root of your project. The skill, per assistant says which ZIP is for which assistant.
• For Claude Code, the skill is also packaged as a plugin.

To work from a clone of this repository instead, build the extension as docs/development.md describes. You need Python 3.10 or newer with the packages of requirements-dev.txt in a virtual environment, and Node 20.18.1 or newer. Then copy the skill into your project with the installer:

python3 tools/install_skill.py /path/to/your/project

The installer copies the skill into the project's .agents/skills/mlview/ folder. For Claude Code, add --destination .claude/skills/mlview.

Use one skill location in a project, not both. The user guide has the details.

A first diagram

1. Open your project in VS Code and start the mlview skill in your assistant. The user guide says how. Ask, for example: "Explain training with train.py and configs/distill.json. Show where data enters, which model is updated, how the losses combine, how validation works, and what source inspection cannot establish. Tell me what looks wrong."
2. The assistant publishes a *.mlview.json file. Click it in the Explorer and it opens as the diagram. Select a step, connection or finding to read its claim and quotes. Press Enter to open the cited lines beside the diagram.
3. Ask the same assistant to refine the diagram. A valid new revision replaces the one on screen.

Example

samples/configured_training.mlview.json is a diagram of the small project in samples/configured_training/. Open this repository in VS Code with the extension installed and click the file. The file records which assistant and model wrote it.

What MLView checks, and what it does not

• The helper and the extension check a document's structure, that every cited path is inside the workspace, the file hashes and the exact quotes.
• A matching quote shows that the cited lines exist and have not changed. It does not show that they support the claim.
• No automated check calls a model or opens VS Code. What is checked, and what is not says what the checks cover.
• Treat a diagram as the assistant's reading of the code, and check it against the cited lines.

In this repository

Path	What it holds
skills/mlview/	The skill: its instructions, reference files and helper, with the helper's tests
contracts/	The JSON Schema of the format, and the cases the schema, the helper and the extension must agree on
vscode-extension/	The VS Code extension
webview/	The diagram viewer the extension shows
claude-plugin/	The Claude Code plugin, with a generated copy of the skill
samples/	Small example projects and example diagrams
tools/, scripts/	Install, package, build and check tools, with their tests
docs/	The user guide, the format reference, and pages on the architecture, development, releasing and what is checked

Develop

Use a Python 3.10+ virtual environment and Node 20.18.1 or newer.

python -m pip install -r requirements-dev.txt
sh scripts/e2e.sh

On Windows run pwsh -File scripts/e2e.ps1. The full check builds the viewer and the extension, type-checks and tests them, runs the Python tests and the documentation link check, and packages the two skill ZIPs and the VSIX. CONTRIBUTING.md has the steps for a change, and docs/development.md has the commands for each part.

More

• The user guide
• The format and the helper's commands, and the format's JSON Schema
• The architecture, development and what is checked, and what is not
• The skill: skills/mlview/SKILL.md
• Contributing, the security policy and the code of conduct
• The changelog, the license and the third-party notices