feat: add Dataflow Gen2 data visuals skill for Claude and ChatGPT

.github/copilot-instructions.md

@@ -28,6 +28,7 @@ When these topics come up, reference the corresponding file in `docs/datafactory
 | Query timeout, slow performance, chunking | `performance.md` |
 | Output destinations, new tables, staging | `destinations.md` |
 | Fast Copy, Action.Sequence, Modern Evaluator | `advanced.md` |
+| Data visuals, dashboards, reports, charts, KPI cards | `data-visuals.md` |
 
 For connection management (creating/listing connections, supported types, UI form), reference `docs/connection-management.md`.
 

.github/skills/datafactory.architecture/SKILL.md

@@ -26,6 +26,7 @@ DataFactory.MCP.Http/        — HTTP transport layer (AspNetCore)
 DataFactory.MCP.Tests/       — Test suite (xUnit)
 DataFactory.WindowsMCP/      — Windows-specific MCP implementation
 claude-skills/               — Claude skill definitions (RAG pattern)
+chatgpt-skills/              — ChatGPT instructions and knowledge files
 docs/                        — Feature documentation
 evals/                       — Evaluation test scenarios
 ```
@@ -89,3 +90,4 @@ For Data Factory / M language / Dataflow patterns, see `claude-skills/`:
 - `datafactory-destinations.md` — Output destinations, DataDestination patterns
 - `datafactory-performance.md` — Query tuning, chunking, query folding
 - `datafactory-advanced.md` — Fast Copy, Action.Sequence, Modern Evaluator
+- `datafactory-data-visuals.md` — Dataflow Gen2 Visual generation and validation

.github/workflows/pr-validation.yml

@@ -40,7 +40,6 @@ jobs:
         uses: actions/setup-dotnet@v4
         with:
           dotnet-version: "10.0.x"
-          include-prerelease: true
 
       - name: 📦 Restore Dependencies
         run: dotnet restore --configfile ../nuget.config
@@ -119,6 +118,12 @@ jobs:
 
             [View full results](${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }})`;
 
+            await core.summary.addRaw(body).write();
+
+            if (context.payload.pull_request.head.repo.full_name !== context.payload.pull_request.base.repo.full_name) {
+              return;
+            }
+
             // Find existing comment
             const comments = await github.rest.issues.listComments({
               owner: context.repo.owner,

.gitignore

@@ -9,3 +9,5 @@ tmpclaude-*
 **/McpApps/node_modules/
 **/McpApps/dist/
 **/build-output.txt
+**/__pycache__/
+*.pyc

README.md

@@ -150,6 +150,7 @@ Upload the skill files from the [`claude-skills/`](claude-skills/) folder to you
    - `datafactory-performance.md` - Query optimization, timeouts, chunking
    - `datafactory-destinations.md` - Output configuration, programmatic setup
    - `datafactory-advanced.md` - Fast Copy, Action.Sequence, Modern Evaluator
+   - `datafactory-data-visuals.md` - Dataflow Gen2 Visual generation
 
 ### What's Covered
 
@@ -159,6 +160,7 @@ Upload the skill files from the [`claude-skills/`](claude-skills/) folder to you
 | **Performance** | Query timeouts, chunking strategies, filter optimization, connector selection |
 | **Destinations** | Lakehouse architecture, schema settings, programmatic destination configuration |
 | **Advanced** | `Action.Sequence` for writes, Fast Copy, Modern Evaluator |
+| **Data visuals** | Visualization documents, visual selection, validation, and limitations |
 
 ### Usage
 
@@ -180,9 +182,27 @@ Create a Custom GPT or use ChatGPT Projects with pre-built Data Factory knowledg
    - `knowledge-performance.md` - Query optimization, timeouts
    - `knowledge-destinations.md` - Output configuration
    - `knowledge-advanced.md` - Fast Copy, Action.Sequence
+   - `knowledge-data-visuals.md` - Dataflow Gen2 Visual contract and workflow
 
 See [`chatgpt-skills/README.md`](chatgpt-skills/README.md) for detailed setup options.
 
+## Dataflow Gen2 Visual Builder (Optional)
+
+Generate Dataflow Gen2 data visuals in Power Query M. **Visual** is the canonical
+artifact name; dashboard, report, and chart are discovery synonyms.
+
+| Client | Knowledge file |
+|--------|----------------|
+| Claude | [`claude-skills/datafactory-data-visuals.md`](claude-skills/datafactory-data-visuals.md) |
+| ChatGPT | [`chatgpt-skills/knowledge-data-visuals.md`](chatgpt-skills/knowledge-data-visuals.md) |
+| Repo agents | [`docs/datafactory/data-visuals.md`](docs/datafactory/data-visuals.md) |
+
+To confirm both clients produce equivalent results, run the checklist in
+[`evals/dataflow-visuals-parity.md`](evals/dataflow-visuals-parity.md).
+
+These are client-side companion files. They are not included in the MCP NuGet
+package, because installing an MCP server does not register skills with a client.
+
 ## Documentation
 
 For complete documentation, see our **[Documentation Index](https://github.com/microsoft/DataFactory.MCP/blob/main/docs/index.md)**.

chatgpt-skills/README.md

@@ -16,6 +16,7 @@ Operational tips and best practices for working with Microsoft Fabric Data Facto
    - `knowledge-performance.md`
    - `knowledge-destinations.md`
    - `knowledge-advanced.md`
+   - `knowledge-data-visuals.md`
 4. Under **Capabilities**, enable:
    - ✅ Code Interpreter (for M code analysis)
 5. Click **Create** → **Save**
@@ -41,6 +42,7 @@ Operational tips and best practices for working with Microsoft Fabric Data Facto
 | `knowledge-performance.md` | Query optimization, timeouts, chunking |
 | `knowledge-destinations.md` | Output configuration, programmatic setup |
 | `knowledge-advanced.md` | Fast Copy, Action.Sequence, Modern Evaluator |
+| `knowledge-data-visuals.md` | Dataflow Gen2 Visual contract, workflow, and limitations |
 
 ## What's Covered
 
@@ -66,6 +68,11 @@ Operational tips and best practices for working with Microsoft Fabric Data Facto
 - Fast Copy (limited transforms, fast ingestion)
 - Modern Evaluator (complex transforms, limited connectors)
 
+### Data visuals
+- Visualization document contract (flat five-column table)
+- Closed PartType set and required properties
+- Visual selection, validation, and Preview limitations
+
 ## Usage Examples
 
 Once your GPT is created, you can ask:
@@ -75,6 +82,7 @@ Once your GPT is created, you can ask:
 - "What's the difference between Fast Copy and Modern Evaluator?"
 - "Help me write an M query to aggregate sales by month"
 - "Explain Action.Sequence and when to use it"
+- "Build a dashboard from my SalesData query"
 
 ## Comparison with Claude Skills
 
@@ -83,3 +91,13 @@ Once your GPT is created, you can ask:
 | Format | Multiple .md files with YAML frontmatter | Instructions + Knowledge files |
 | Loading | On-demand via RAG triggers | All knowledge available |
 | Best for | Claude Projects | Custom GPTs or ChatGPT Projects |
+
+## Dataflow Gen2 Visual builder
+
+Visual requests are routed by `gpt-instructions.md`, which loads
+`knowledge-data-visuals.md`. Visual is the canonical artifact name; dashboard,
+report, and chart remain discovery synonyms.
+
+Use the shared cases in `../evals/dataflow-visuals-parity.md` to run the same
+prompts in ChatGPT and Claude and compare both outputs with one checklist. No
+provider API keys are required.

chatgpt-skills/gpt-instructions.md

@@ -50,6 +50,14 @@ You are an expert assistant for Microsoft Fabric Data Factory, specializing in M
 4. Offer to open the interactive form: resource `ui://datafactory/create-connection`
 5. Use `list_connections` / `get_connection` to inspect existing connections
 
+### When users ask for a visual, dashboard, report, chart, or KPI:
+1. Load `knowledge-data-visuals.md`
+2. Call the artifact a Visual or visualization document
+3. Clarify business intent and ambiguous metric definitions
+4. Generate the documented flat five-column Power Query M table
+5. Validate hierarchy, property types, and exact chart-column mappings
+6. Ask before persistent changes and validate the complete mashup before saving
+
 ## Response Style
 
 - Be concise and practical
@@ -78,3 +86,4 @@ in
 - Fast Copy only supports: combine files, select columns, change types, rename/remove columns
 - For connection creation, always call `list_supported_connection_types` first to get correct parameters
 - `gatewayId` is required for `OnPremisesGateway` and `VirtualNetworkGateway` connectivity types
+- Data visuals are a Preview feature; structural validation does not prove successful Fabric rendering

chatgpt-skills/knowledge-data-visuals.md

@@ -0,0 +1,56 @@
+# Dataflow Gen2 Data Visuals
+
+Use this knowledge for Dataflow Gen2 visual, visualization, dashboard, report,
+chart, KPI, and visual-summary requests. **Visual** and **visualization document**
+are canonical; the other terms are request synonyms.
+
+Data visuals are in Preview. Return one flat Power Query M table whose exact
+columns are `Name` (`nullable text`), `Parent` (`nullable text`), `PartType`
+(`nullable text`), `Properties` (`nullable record`), and `Data` (`any`).
+
+The closed PartType set is `Container`, `Card`, `Header`, `KpiCard`, `Table`,
+`LineChart`, `AreaChart`, `BarChart`, `StackedBarChart`, `DonutChart`, and
+`PieChart`. Never invent another value or use a nested record-tree contract.
+
+## Contract
+
+| PartType | Children | Required values |
+| --- | --- | --- |
+| `Container` | One or more | Optional `Direction` is `"row"` or `"column"` |
+| `Card` | Exactly one | Text `Title` |
+| `Header` | None | Text `Header`; optional text `FarText` |
+| `KpiCard` | None | Text `Value`, text `Label`; optional text `Sub` |
+| `Table` | None | Table in `Data` |
+| `LineChart`, `AreaChart` | None | `XAxis`, numeric `YAxis`, table in `Data` |
+| `BarChart`, `DonutChart`, `PieChart` | None | `Category`, numeric `Value`, table in `Data` |
+| `StackedBarChart` | None | `Category`, numeric `Value`, `Series`, table in `Data` |
+
+Exactly one row has `Parent = null`. Use unique non-null names, resolvable
+parents, and acyclic relationships. Charts go inside cards. Chart properties
+name exact columns in `Data`; a mismatch can silently render an `undefined`
+bucket. Format KPI values as text.
+
+Duplicate names don't raise an error, but any row parented to an ambiguous name
+fails to render, so keep names unique.
+
+## Workflow
+
+1. Clarify objective, audience, measures, dimensions, time grain, filters,
+   targets, and exclusions. Confirm ambiguous definitions.
+2. Inspect query definitions or metadata without sampling business values.
+   Prefer `get_dataflow_definition` when MCP tools are available. Never use
+   `Expression.Evaluate` or dynamically evaluate sibling query values.
+3. Recommend a lightweight Visual and state assumptions.
+4. Select required columns, filter invalid values, aggregate to visual grain,
+   sort explicitly, and cap categories before constructing the flat `#table`.
+5. Begin with at most three data visuals, ten chart categories, and 50 detail
+   rows. Add complexity only after rendering succeeds.
+6. Test with `execute_query` when possible. Ask before persistent changes. Use
+   `save_dataflow_definition` with `validateOnly = true` before saving because
+   a save replaces the complete mashup. After saving, restore required bindings
+   with `add_connection_to_dataflow` and validate them.
+
+Visuals are static, appear only in the authoring canvas, do not become refresh
+output, and can slow authoring when numerous or backed by large tables. Line and
+area charts do not fill missing dates. Structural validation does not prove a
+successful Fabric render.

claude-skills/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: datafactory-tips
-description: Operational tips and best practices for working with Microsoft Fabric Data Factory MCP tools. Use when executing queries, troubleshooting timeouts, creating dataflows end-to-end, or working with large datasets via the DataFactory.MCP tools.
+description: Operational guidance for Microsoft Fabric Data Factory MCP tools. Use when executing queries, troubleshooting timeouts, creating dataflows, or building a Dataflow Gen2 visual, dashboard, report, chart, or KPI.
 ---
 
 # Data Factory Tips
@@ -39,6 +39,10 @@ Operational knowledge for working with Microsoft Fabric Data Factory.
 | Multi-source instant fail via API | Dirty dataflow or separate Lakehouse.Contents calls | `sources/multi-source.md` |
 | `IsNewTarget = false` fails | Direct navigation on API-created dataflow | `destinations/dest-new-table.md` |
 | Stale connections after revert | save_dataflow_definition doesn't remove connections | `datafactory-connections.md` (Troubleshooting) |
+| Visual renders as a plain table | A required column is missing or renamed | `datafactory-data-visuals.md` |
+| `Visual not recognized: "<value>"` | PartType outside the closed set | `datafactory-data-visuals.md` |
+| Chart shows one `undefined` bucket | Property names a column missing from `Data` | `datafactory-data-visuals.md` |
+| `must contain exactly one root row` | Zero or multiple rows with `Parent = null` | `datafactory-data-visuals.md` |
 
 ## Knowledge Files
 
@@ -49,6 +53,7 @@ Operational knowledge for working with Microsoft Fabric Data Factory.
 | `datafactory-performance.md` | Query timeouts, chunking, query folding, connector selection |
 | `datafactory-advanced.md` | Fast Copy limits, Action.Sequence, Modern Evaluator |
 | `datafactory-pipelines.md` | Pipeline creation, Dataflow activities, chaining, scheduling |
+| `datafactory-data-visuals.md` | Dataflow Gen2 Visual, dashboard, report, chart, and KPI generation |
 
 ### Destination Files (read only the one you need)
 

claude-skills/datafactory-data-visuals.md

@@ -0,0 +1,66 @@
+# Dataflow Gen2 Data Visuals
+
+Use this knowledge when a user asks for a Dataflow Gen2 visual, visualization,
+dashboard, report, chart, KPI, or visual summary. Call the generated artifact a
+**Visual** or **visualization document**; the other terms are request synonyms.
+
+Data visuals are a Preview feature. A Visual is a flat Power Query M table with
+these exact columns and types:
+
+```powerquery-m
+type table [
+    Name = nullable text,
+    Parent = nullable text,
+    PartType = nullable text,
+    Properties = nullable record,
+    Data = any
+]
+```
+
+Only these PartTypes exist: `Container`, `Card`, `Header`, `KpiCard`, `Table`,
+`LineChart`, `AreaChart`, `BarChart`, `StackedBarChart`, `DonutChart`, and
+`PieChart`.
+
+## Contract
+
+| PartType | Children | Required values |
+| --- | --- | --- |
+| `Container` | One or more | Optional `Direction` is `"row"` or `"column"` |
+| `Card` | Exactly one | Text `Title` |
+| `Header` | None | Text `Header`; optional text `FarText` |
+| `KpiCard` | None | Text `Value`, text `Label`; optional text `Sub` |
+| `Table` | None | Table in `Data` |
+| `LineChart`, `AreaChart` | None | `XAxis`, numeric `YAxis`, table in `Data` |
+| `BarChart`, `DonutChart`, `PieChart` | None | `Category`, numeric `Value`, table in `Data` |
+| `StackedBarChart` | None | `Category`, numeric `Value`, `Series`, table in `Data` |
+
+Exactly one row has `Parent = null`. Every `Name` is unique and non-null. Every
+other `Parent` resolves to a `Name`, and relationships are acyclic. Charts go
+inside cards. Chart properties contain exact column names from their `Data`
+table; a bad name can silently produce one `undefined` bucket. KPI values are
+formatted as text.
+
+Duplicate names don't raise an error, but any row parented to an ambiguous name
+fails to render, so keep names unique.
+
+## Workflow
+
+1. Clarify objective, audience, measures, dimensions, time grain, filters,
+   targets, and exclusions. Confirm ambiguous definitions.
+2. Inspect query definitions or metadata without sampling business values.
+   Prefer `get_dataflow_definition` when MCP tools are available. Never use
+   `Expression.Evaluate` or dynamically evaluate sibling query values.
+3. Recommend a lightweight Visual and state assumptions.
+4. Select required columns, filter invalid values, aggregate to visual grain,
+   sort explicitly, and cap categories before building the flat `#table`.
+5. Start with at most three data visuals, ten chart categories, and 50 detail
+   rows. Add complexity only after rendering succeeds.
+6. Test nonpersistently with `execute_query` when possible. Ask before changing
+   a dataflow. Use `save_dataflow_definition` with `validateOnly = true` before
+   saving because a save replaces the complete mashup. After saving, restore
+   required bindings with `add_connection_to_dataflow` and validate them.
+
+Visuals are static, render only in the authoring canvas, do not become refresh
+output, and can slow authoring when numerous or backed by large tables. Line and
+area charts do not fill missing dates. Structural validation cannot prove that
+Fabric rendered the Visual successfully.