Semantic Model Cleaner for VS Code - Preview

Choose a semantic model and its reports in a native Cleaner sidebar, then review the analysis in a desktop VS Code editor tab running the existing Semantic Model Cleaner web UI. The Python engine discovers the Power BI Project and its report bindings. Analysis and model changes remain explicit user actions; the existing reviewed-plan safeguards are unchanged.

Version 0.1.7 is prepared as a Marketplace pre-release for MrPerfectH.semantic-model-cleaner; this document does not confirm publication. The Preview label remains. The VSIX contains the extension, not Python or the engine. Nothing is downloaded, installed or upgraded automatically.

Requirements and installation

• Desktop VS Code 1.95 or newer and a trusted local workspace.
• Python 3.11+ and the separately installed Semantic Model Cleaner 0.4.0b10 engine from the same reviewed build.
• Local disk Power BI Project files (TMDL Semantic Model and PBIR Reports).

Use the separately prepared semantic_model_cleaner-0.4.0b10-py3-none-any.whl from the same reviewed build; this document does not confirm beta10 publication or provide a public beta10 download URL. Place that wheel in the current directory before running the commands below. Install it into a virtual environment at a location you choose. These commands create an example environment in the current directory; run them yourself in a terminal.

POSIX shell (macOS/Linux):

python3 -m venv .venv-smc
.venv-smc/bin/python -m pip install "./semantic_model_cleaner-0.4.0b10-py3-none-any.whl"

Windows PowerShell:

py -3 -m venv .venv-smc
& .\.venv-smc\Scripts\python.exe -m pip install "./semantic_model_cleaner-0.4.0b10-py3-none-any.whl"

Install the extension through Extensions: Install from VSIX... using semantic-model-cleaner-0.1.7.vsix. Once the public pre-release is available, its Marketplace identity is MrPerfectH.semantic-model-cleaner. Set Semantic Model Cleaner: Python Path to the absolute path of your chosen environment's Python executable (.venv-smc/bin/python or .venv-smc\Scripts\python.exe). Alternatively leave Python Path empty and set Executable to that environment's absolute smc-web / smc-web.exe path. The extension launches that configured external engine; installing this VSIX does not replace or configure an existing engine or executable setting.

Settings are resource-scoped and restricted by Workspace Trust. The extension also checks trust in command handlers before reading executable settings or launching a process. Values are single executable paths or PATH command names, never shell commands. Quotes, interpreter arguments, variable substitution and .cmd/.bat wrappers are not supported. Paths containing spaces work without adding quotes. Python mode uses -I -m semantic_model_cleaner.web to avoid importing Python modules from the selected project directory.

Use

Select the Cleaner icon in the Activity Bar. Analysis Scope is the native picker; Cleaner Actions (collapsed) keeps Open Cleaner, Restart Server, Stop Server and Show Logs. All use the same trust-checked commands as the Command Palette.

• Open Cleaner (in Analysis Scope) uses the single workspace folder, offers a picker in multi-root workspaces, or a folder dialog in an empty window. Right-click a folder in Explorer to open it directly. Nothing starts until you choose it.
• Cold commands: Choose Semantic Model, Analyze and report selection commands open Cleaner and wait for model/report discovery when necessary. Open reveals the native Analysis Scope sidebar. Open and Choose never run analysis or Apply. Canceling a folder/model picker or stopping startup abandons the pending action.
• Semantic model: a sole model is selected automatically. With several, select one explicitly; nothing is guessed. Unsupported model folders are listed with a warning.
• Reports: every report in the folder connected to the chosen model is selected by default (3 of 5 selected). Clear checkboxes to analyze a subset, or use Select All Reports / Clear Report Selection. Choosing another model resets to all its connected reports. Reports whose binding cannot be verified are shown but not selectable; reports of other models are counted.
• Analyze analyzes exactly the chosen model and reports. With no reports it inspects the model only. A subset or zero reports leaves usage in the other reports unknown, so the result records an analysis limitation and no item is called Safe.
• The editor shows results only; it has no scope drawer. When the sidebar changes after an analysis, the editor marks the results out of date and blocks actions on them until you analyze again. A change during an analysis discards that run's result.
• Restart Server reloads settings and restarts the last selected project, keeping its model and report choices. Opening another folder resets them.
• Stop Server closes the panel and stops its process. Closing the panel or deactivating the extension also stops the process.
• Show Logs displays bounded, token-redacted launch and server output. The status bar shows starting/running/stopped/error and opens Cleaner.

One panel and one owned server run per VS Code window. Opening another folder replaces the current one. Scope commands wait for the current startup and discovery. Opening another folder replaces the previous launch; a stopped or closed panel cannot start a late server. No existing server on a guessed port is reused.

Boundaries

• Local desktop only. Remote SSH, WSL, Dev Containers, Codespaces, virtual workspaces and VS Code in a browser fail before starting a process. A workspace extension and loopback portMapping are used, but remote operation is not claimed. Public tunnels and asExternalUri URLs would need separate Host, Origin and ancestor verification; this preview deliberately does not create one.
• The engine binds 127.0.0.1 on an OS-assigned port. A fresh 256-bit launch token travels only in the child's environment and request headers. Readiness checks the matching instance fingerprint. There are no token URLs or token files.
• An opt-in --vscode --host 127.0.0.1 --port 0 protocol permits only the exact desktop webview and shell ancestor origins measured from the panel. Only the root HTML iframe navigation is exempt from the normal Fetch Metadata framing rejection. Host, peer, API token, JSON and Origin checks remain enforced; there is no CORS permission. Ordinary web launches still deny embedding.
• The panel has no local resource roots and can frame only its own loopback server. The existing UI uses local assets; this extension adds no CDN requests.
• VS Code's sandbox blocks native alert/confirm dialogs. Embedded mode uses in-page alerts and awaited confirmations; Cancel has initial focus and Escape cancels. Ordinary browser mode retains native dialogs. Existing app-local browsing and explicitly confirmed plan actions still operate on the local filesystem; the selected folder is a starting scope, not a filesystem jail.
• Logs are limited to 64 KiB. Secret redaction is applied to complete bounded lines, so a token split across stdout chunks cannot leak. Overlong or trailing incomplete lines are omitted. An extension-host crash closes the child's stdin lifetime pipe; ordinary stop also signals and, if needed, force-kills the child.

Develop, test and package

From this directory, with Node 22+:

npm ci
npm test
npm run package

The package command creates semantic-model-cleaner-0.1.7.vsix with the VSIX pre-release flag. It does not publish. Open this directory in VS Code and use Run Cleaner extension (F5) for a development host.

Run the Python suite from the repository root:

.venv/bin/python -m pytest
.venv/bin/python -m ruff check .

A real extension-host/browser smoke is available:

SMC_VSCODE_EXECUTABLE='/Applications/Visual Studio Code.app/Contents/MacOS/Code' npm run test:integration

SMC_TEST_PYTHON can select the prepared interpreter. Without SMC_VSCODE_EXECUTABLE, the VS Code test runner downloads its default test host. Linux needs a desktop or xvfb-run -a npm run test:integration. Windows can set these environment variables in PowerShell. The script builds a disposable two-model project (bundled demo plus the Product QA example, extra connected reports and one broken binding), uses disposable VS Code/SMC profiles and a local debugging port, and deletes the copy afterward. Set SMC_TEST_VSIX to install and test a packaged VSIX in the disposable extensions directory instead of this source folder (Windows also needs SMC_VSCODE_CLI_JS, the product's resources/app/out/cli.js). It disables Workspace Trust only in that disposable test host; trust denial is covered separately by command-handler tests. It drives the native sidebar: cold Choose and Analyze, explicit model choice, default and deselected reports, exact subset and model-only analysis, stale results, model and folder switches, a change during analysis, restart and stop, and hashes inputs to detect changes. Evidence goes to integration-artifacts/. No user project is used.

See ACCEPTANCE.md for observed results, evidence boundaries and installed-VSIX verification steps.

API references: Workspace Trust, webview security, and desktop port mapping.

Marketplace release preparation

The package declares the public publisher MrPerfectH, repository URL and MIT license, and includes LICENSE. The media/cleaner-broom.svg is the Activity Bar icon; media/cleaner.png is the Marketplace package icon.

The manually dispatched Publish Cleaner Marketplace pre-release workflow runs extension and focused Python checks, packages this version and publishes that exact VSIX with --pre-release. A repository maintainer must configure VSCE_PAT with permission for the MrPerfectH publisher and explicitly dispatch the workflow from the reviewed release commit. Pushes do not publish. Packaging or passing tests alone does not establish Marketplace availability.