MLView

Use the LLM in your existing VS Code assistant to understand an ML workflow, then explore its source-linked interactive diagram. MLView provides a skill for GitHub Copilot, Codex, and Claude Code, a common artifact format, and a VS Code viewer. Your assistant reads the source and configuration, interprets the workflow, and authors the diagram and findings.

MLView requires no separate model API key. The selected assistant supplies the model, tools, permissions, and source-processing policy. Local helpers validate citations and structure; the viewer displays the result. Neither helper nor viewer executes the analyzed program. Citation validation does not prove the model's interpretation correct.

This is an experimental implementation. See implementation and validation status for what has actually been tested. Workflow interpretation and findings come from the active assistant model.

Get a diagram

The detailed install and refinement instructions are in LLM_WORKFLOW.md. From this checkout, with Python 3.10+ and Node 20.18.1+:

cd webview && npm ci && npm run build && cd ..
python3 tools/sync-assets.py
cd vscode-extension && npm ci && npm run package && cd ..
python3 tools/install_skill.py /path/to/your/project

Install the generated VSIX using VS Code's Extensions: Install from VSIX. The installer copies the self-contained skill into the target project's .agents/skills/mlview/ directory for Codex and Copilot. For Claude Code, use the Claude plugin or the workspace install described in the runbook. Only install one copy of the skill in each discovery path.

1. Open the target project in VS Code and invoke mlview in your assistant's skill picker ($mlview in Codex). Ask, for example: “Explain training with this config. Show data, model, losses, parameter updates, validation, and anything unresolved.”
2. The assistant publishes a *.mlview.json file. Click it in the Explorer: it opens as the diagram (to see the JSON, use View: Reopen Editor With… → Text Editor). MLView: Open Generated Diagram opens it beside the editor you are in (or brings its open diagram to the front of its own group). Click a step, connection or finding to see its claim and quotes. Press Enter, double-click, or use an Open link to open the cited lines beside the diagram; focus stays in the diagram. Alt+Enter opens them and moves focus to the editor.
3. Ask the same assistant to refine the diagram. A valid new revision updates the panel; filtering and navigating the existing diagram do not call an LLM.

To refine a particular node, edge, or finding, select it and use Refine → Copy prompt. Choose an intent or enter a specific question, then paste the prompt into the same assistant. The header is one row: the title, the assistant and revision, search, the severity toggles, N not observed, Review, a ⋯ menu (legend, flow animation, fit, exports, shortcuts) and Refine…. In a narrow panel, such as beside your code, search, the revision and Review move into the ⋯ menu, and a row that is still too full folds N not observed into it too, so Refine… stays in view. The status bar counts steps and connections, shows the coverage status with its limitations, and says whether the cited files are unchanged (in muted text) or changed (with a warning).

Beside the diagram, the rail has four tabs: About, Findings, Selection and Outline. A new revision opens on About: the question that was asked, what the model says it traced, the coverage with its limitations (listed once), the scope, the run configuration, the cited files with their freshness, and who wrote it ("Model-authored; MLView checks citations, not the interpretation"). Selection shows the selected claim first: its phase, title and full detail, a short note when it is inferred or unresolved, the findings on it with What to change, its numbered source quotes with line numbers and Open, and what it comes from and feeds (for a group: the findings inside it, its steps and the connections across its edge). A click on the diagram or a search hit shows the claim in Selection; a row chosen in the Findings list or the Outline keeps that list in place. A matching quote only shows that the cited lines are unchanged since publishing; whether they support the claim is for you to judge. Challenge this claim prepares a focused refinement request. Selecting a finding frames every step it cites. The Outline's textual relationships provide All, Incoming, Outgoing and Unresolved views alongside the diagram. In a panel narrower than about 1260 px the rail is a bottom sheet under the diagram: a tab strip until you select something, then about half the height, with the selected card kept in view above it; Escape collapses it.

To check a diagram claim by claim, press r (or Review). The review walk starts on the claims marked inferred or unresolved, the header's N not observed count, and can also walk the findings, every claim, or, after a cited file changed, the claims citing it. Each step selects the claim, shows it in Selection and, after a short pause, opens its cited lines beside the diagram, highlighted, while the keyboard stays on the diagram: j / k (or ↓ / ↑, or the walk bar's up and down buttons) move, [ / ] change quote, Enter opens again and Escape ends the walk. A quote whose file changed is not opened; the walk says why. The walk remembers its place for each revision and records nothing else.

When the assistant publishes a refined revision whose parent is the one the panel is showing, About opens on Changes since that revision: the steps, connections and findings added, removed or changed, matched by id. Added and changed items are links (a changed one says what changed, such as the label or the evidence); removed ones are listed by name and never drawn. Added and changed steps carry a small "new" or "changed" tag on their card, connections and findings carry it in the Outline, the Findings list and Selection, and the walk gains Changed in this revision. A renamed phase is listed once and marks none of its steps. Only revisions this panel has shown are compared: closing the panel or reloading the window forgets the comparison (a diagram in a preview tab is kept open once it shows changes, so another file does not replace it). A tag says the authored text changed, not whether the claim holds.

To see the whole workflow at once, press Shift+0 (or Phase overview in the ⋯ menu). The overview shows each phase as a block of its step titles, with inferred and unresolved marks and the F labels of findings. Arrows count the connections from one phase to the next, and brackets count those that skip ahead or go back. Arrow keys and Enter (or a click) go to a phase; Escape returns you to where you were. The phase index in the diagram's lower right corner lists every phase with its findings and step count and marks the ones in view. Beside the code it shrinks to one line such as "4/6 Objective, optimizer & scheduler".

To go the other way, from the code to the diagram, right-click a line in the editor and choose Reveal in Diagram (also MLView: Reveal in Diagram in the Command Palette; there is no default keybinding). It is offered on a file an open diagram cites that has not changed since the revision was published. The claim citing that line, or the selected lines, is selected and shown in the diagram, and the keyboard moves there. When several claims cite the line, a list asks which; on a line no claim cites, it offers the nearest claims in the file. It works in notebook cells too.

The workflow supports custom phases, nested groups, branches and cycles, notebook cell references, and findings with evidence and counter-evidence. Observed, inferred, and unresolved claims remain distinguishable: only the inferred (dashed, with an inferred tag) and unresolved (dotted, with a ? unresolved tag) ones are marked on the diagram, and the header's N not observed toggle fades the rest. Each phase is coloured by its place in the document; colour is otherwise used only for problems. Finding badges read F1, F2… in document order; the Selection tab and Refine prompts keep the real finding id. The diagram opens at a zoom where cards can be read: the whole document when it fits at 75% or more, otherwise the first phase at 90%, so card titles open at about 10 px or larger, and a wider panel never opens them smaller unless it shows the whole document. Press 0 to return to that view, or use Fit the whole diagram in the ⋯ menu to see everything; zoomed out, cards show just their titles at about 11 px. Saved source edits mark the steps, connections, findings and quotes that cite the changed files, and block jumps into those files; malformed updates retain the last valid diagram. When the workspace root is a parent of the folder the diagram cites from, the viewer says so and offers to add that folder to the workspace.

How it works

Try the native Codex-generated example from this repository's workspace. Its validation log records native model invocations, citation checks, live navigation/refinement and remaining acceptance checks.

flowchart LR
    A[Native 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

The portable skil