Microsoft AI-first Development Weekly: deployment checks
Microsoft AI-first Development Weekly covers Power Apps Code Apps, Microsoft Copilot Managed Runtime and Power Platform skills and plugins for AI-first development.
A note on Copilot Managed Runtime: its command-line package documentation refers to a Preview database, but we have not found a formal Microsoft announcement explaining what it is. It is also unclear whether the deployment migrations described below concern that same store. These terms appear in Microsoft's package README; the database itself remains unexplained.
Copilot Managed Runtime deployment waits for data migrations
The Copilot Managed Runtime command-line tool now waits for an accepted data migration before it reports a deployment as successful. If the wait reaches its limit, it returns the operation ID and an error while the deployment continues, so the command is not silently resubmitted. Makers can check the live commit with ms app info. The published package README describes the behaviour. Microsoft Learn still labels Copilot Managed Runtime a preview; access and runtime entitlement depend on the tenant and user.
Code Apps connection setup is more flexible
The Power Apps Code Apps command-line tool can now add several tables from one connector, connection and dataset in a single command. It reports successes and failures separately if one table fails. For connectors such as SharePoint that offer both tables and actions, --as table or --as action makes the choice explicit, which matters in non-interactive automation. Interactive use can also prompt for connections, datasets and tables. Native Dataverse tables and SQL stored procedures still require one item per command. Microsoft's connection guide and command reference document the new behaviour.
Model Apps plugin protects approved changes
The Model Apps agent plugin now binds destructive approval to the removals the maker was shown. If another field or navigation target would be removed before the approved rerun, the build stops instead of treating the old approval as permission. It also avoids asking for destructive approval when prune: false means a form field will be kept. Microsoft's change record includes before-and-after checks.
The plugin also identifies dashboards by their owning solution when names collide. Teardown keeps dashboards and other resources that the app does not own, and refuses deletion if it cannot record its safety fence. This reduces the risk of a rebuild or teardown changing another app's work. The dashboard and teardown changes are separate from the approval fix.
Model Apps build results are easier to trust
The plugin corrected how it writes per-app AI settings. An enabled natural-language grid search or Microsoft 365 Copilot setting could previously be stored as Off because the settings use different numeric meanings. Build and verification now use each setting's own values. The AI settings fix explains the previous and current behaviour.
Form builds now move tabs and sections into the specified order, narrow fields that would overflow a section, and verify visibility, width and read-only state. The same update keeps a generative page's name and model when an upload omits them. These checks make a successful verification more useful when editing existing apps. The form and page update details the changes.
The generative-page workflow also checks that the approved plan targets the intended files and that output is complete before upload. It rejects unsafe paths and stale files from an earlier attempt. That matters when an agent generates or edits pages in a shared workspace. The page safety update records the new gates. Transpiling a generative page converts its code but does not check its types.
Model Apps rebuilds have important limits
Microsoft added explicit limits to the Model Apps plugin's rebuild guidance. A downloaded app specification carries a generative page's code and data sources, but not its connector or custom API bindings. A rebuild in the same environment preserves existing bindings; a rebuild into another environment does not recreate them. Those bindings need to be configured in the target environment through /genpage.
Existing views also use additive updates: removing a column from the specification does not remove it from the deployed view. Verification tolerates extra columns, so a successful check cannot confirm a removal. The revised capability guidance documents these limits.
Experimental data reset described in the command-line package
Microsoft's Copilot Managed Runtime command-line package README describes ms data clear, an experimental command that requests the deletion of what it calls the app's Preview database, for all users and all item types. The README says live data, source code and the app itself are unchanged. It does not explain the underlying storage or how this database is provisioned.
The command is hidden unless MS_CLI_FEATURE_DATA_CLEAR=1 is set. It requires explicit confirmation, and a successful request does not prove the purge has finished. Do not start dependent Preview writes or a deployment while completion remains unconfirmed. Microsoft's package README sets out those limits.
Keep up with AI first development with Power Squared on YouTube.
Comments