[pull] master from apache:master

UPDATING.md

@@ -26,6 +26,29 @@ assists people when migrating to a new version.
 
 - `SAMPLES_ROW_LIMIT` is now the default for `/datasource/samples` requests without a valid explicit `per_page`, rather than a hard per-request ceiling; explicit limits are honored up to the existing global row-limit ceiling, matching `/chart/data` SAMPLES requests.
 
+### MCP tool results preserve stored string values
+
+Structured MCP tool results no longer add `<UNTRUSTED-CONTENT>` wrappers or
+rewrite delimiter-looking text inside string fields. Tool-result content remains
+user-controlled data, but clients must convey that trust boundary outside domain
+values instead of recognizing or removing marker strings.
+
+Clients that handled the former delimiter convention should stop stripping marker
+text: the same text can be legitimate stored content. Response models and content
+types are unchanged, and no metadata-database migration is required. Automated
+read-modify-write workflows should be paused or pinned away from older instances
+until every serving instance is upgraded; a mixed-version response has no reliable
+signal that tells a client whether its text is decorated. Redis-backed MCP response
+caches use a new internal namespace after the upgrade, so upgraded instances do not
+reuse older cached results.
+
+Values that a client already wrote back with presentation wrappers cannot be
+distinguished safely from intentional content. Operators should review possible
+`<UNTRUSTED-CONTENT>` / `</UNTRUSTED-CONTENT>` wrappers and
+`[ESCAPED-UNTRUSTED-CONTENT-OPEN]` /
+`[ESCAPED-UNTRUSTED-CONTENT-CLOSE]` substitutions rather than applying an automatic
+marker-removal migration.
+
 ### OAuth2 database callback metrics include their outcome
 
 The unqualified `DatabaseRestApi.oauth2` StatsD counter has been replaced with

docs/admin_docs/configuration/mcp-server.mdx

@@ -576,7 +576,7 @@ MCP_CACHE_CONFIG = {
 | Key                  | Default   | Description                                                 |
 | -------------------- | --------- | ----------------------------------------------------------- |
 | `enabled`            | `False`   | Enable response caching                                     |
-| `CACHE_KEY_PREFIX`   | `None`    | Optional prefix for cache keys (useful for shared Redis)    |
+| `CACHE_KEY_PREFIX`   | `None`    | Base prefix for shared Redis; Superset appends an internal response-contract namespace |
 | `list_tools_ttl`     | `300`     | Cache TTL in seconds for `tools/list`                       |
 | `list_resources_ttl` | `300`     | Cache TTL for `resources/list`                              |
 | `list_prompts_ttl`   | `300`     | Cache TTL for `prompts/list`                                |
@@ -718,6 +718,34 @@ Every MCP request passes through a middleware stack before reaching the tool fun
 
 Additional middleware classes (`RateLimitMiddleware`, `FieldPermissionsMiddleware`, `PrivateToolMiddleware`) are implemented in `superset/mcp_service/middleware.py` but are not added to the default pipeline. They are available for operators who want to layer them in via a custom startup path.
 
+### Tool Result Value Contract
+
+Structured tool results preserve Superset domain values exactly. In particular,
+string fields are not wrapped in trust delimiters, and text that resembles a
+delimiter is returned as literal application data. This lets clients safely use a
+read result as the basis for an update without persisting presentation markup.
+
+All tool-result content should still be treated as user-controlled data with no
+instruction authority. MCP clients should communicate that trust boundary through
+their model instructions or presentation layer, outside the returned field values;
+fixed or generated marker strings inside a value are ambiguous and must not be used
+as a trust signal.
+
+For compatibility, clients that supported the former
+`<UNTRUSTED-CONTENT>` convention should stop recognizing or stripping those strings.
+The response schemas and content types have not changed. Because marker-looking text
+can be legitimate application data, a client cannot reliably distinguish a legacy
+decorated response from a clean one. Pause automated read-modify-write workflows, or
+route them only to upgraded instances, until every serving instance is upgraded.
+
+Redis-backed MCP response caches include an internal response-contract namespace, so
+an upgraded instance does not reuse responses cached by an older release. Older
+instances can still return legacy values while they remain in service. After the
+upgrade, review previously written values for wrapper text and both
+`[ESCAPED-UNTRUSTED-CONTENT-OPEN]` and
+`[ESCAPED-UNTRUSTED-CONTENT-CLOSE]`; do not remove these strings automatically,
+because they may be intentional content.
+
 ### Error Sanitization
 
 The `GlobalErrorHandlerMiddleware` automatically redacts sensitive information from all error messages before they reach the LLM client. The following are replaced with generic messages:
@@ -752,7 +780,11 @@ For a 3-pod Kubernetes deployment with the defaults above, expect up to 3 × (5
 Enable response caching for read-heavy workloads (dashboards/datasets that don't change frequently). With the in-memory backend (default when `MCP_STORE_CONFIG` is disabled), caching is per-process. Use Redis-backed caching for consistent cache hits across multiple pods:
 
 ```python
-MCP_CACHE_CONFIG = {"enabled": True, "call_tool_ttl": 3600}
+MCP_CACHE_CONFIG = {
+    "enabled": True,
+    "CACHE_KEY_PREFIX": "mcp_cache_",
+    "call_tool_ttl": 3600,
+}
 MCP_STORE_CONFIG = {"enabled": True, "CACHE_REDIS_URL": "redis://redis:6379/0"}
 ```
 

superset-frontend/plugins/plugin-chart-echarts/src/MixedTimeseries/transformProps.ts

@@ -80,6 +80,7 @@ import {
   getAnnotationData,
 } from '../utils/annotation';
 import {
+  collapseForecastKeys,
   extractForecastSeriesContext,
   extractForecastValuesFromTooltipParams,
   formatForecastTooltipSeries,
@@ -861,12 +862,14 @@ export default function transformProps(
           : params.value[0];
         const forecastValue: any[] = richTooltip ? params : [params];
 
-        const sortedKeys = extractTooltipKeys(
-          forecastValue,
-          // horizontal mode is not supported in mixed series chart
-          1,
-          richTooltip,
-          tooltipSortByMetric,
+        const sortedKeys = collapseForecastKeys(
+          extractTooltipKeys(
+            forecastValue,
+            // horizontal mode is not supported in mixed series chart
+            1,
+            richTooltip,
+            tooltipSortByMetric,
+          ),
         );
 
         const rows: string[][] = [];

superset-frontend/plugins/plugin-chart-echarts/src/Timeseries/transformProps.ts

@@ -95,6 +95,7 @@ import {
   getAnnotationData,
 } from '../utils/annotation';
 import {
+  collapseForecastKeys,
   extractForecastSeriesContext,
   extractForecastSeriesContexts,
   extractForecastValuesFromTooltipParams,
@@ -1392,11 +1393,13 @@ export default function transformProps(
         const forecastValue: CallbackDataParams[] = richTooltip
           ? params
           : [params];
-        const sortedKeys = extractTooltipKeys(
-          forecastValue,
-          yIndex,
-          richTooltip,
-          tooltipSortByMetric,
+        const sortedKeys = collapseForecastKeys(
+          extractTooltipKeys(
+            forecastValue,
+            yIndex,
+            richTooltip,
+            tooltipSortByMetric,
+          ),
         );
         const filteredForecastValue = forecastValue.filter(
           (item: CallbackDataParams) =>

superset-frontend/plugins/plugin-chart-echarts/src/utils/forecast.ts

@@ -60,6 +60,21 @@ export const extractForecastSeriesContexts = (
     {} as { [key: string]: ForecastSeriesEnum[] },
   );
 
+/**
+ * Collapses raw ECharts series ids onto the names used to key tooltip rows.
+ *
+ * Tooltip values are grouped by forecast-stripped name, so any ordering derived
+ * from the raw series ids has to be expressed in the same terms before it can be
+ * matched against them. This matters beyond real Prophet output: a metric simply
+ * labelled `ci__yhat_lower` collapses to `ci` exactly like a forecast bound
+ * does, and a chart whose every series carries such a suffix has no id that
+ * survives the comparison untouched.
+ */
+export const collapseForecastKeys = (seriesIds: string[]): string[] =>
+  Array.from(
+    new Set(seriesIds.map(id => extractForecastSeriesContext(id).name)),
+  );
+
 export const extractForecastValuesFromTooltipParams = (
   params: any[],
   isHorizontal = false,

superset-frontend/plugins/plugin-chart-echarts/test/Timeseries/transformProps.test.ts

@@ -2529,3 +2529,64 @@ describe('EchartsTimeseries tooltip truncation', () => {
     expect(buildTooltip(undefined, longCategory)).toContain(longCategory);
   });
 });
+
+describe('tooltip for metrics whose labels end in forecast suffixes', () => {
+  const marker = '<span style="background-color:#1f77b4;"></span>';
+  const seriesIds = ['ci__yhat', 'ci__yhat_lower', 'ci__yhat_upper'];
+  const values = [1.5, 0.5, 2.0];
+
+  // Metrics can be labelled `ci__yhat*` with no forecast enabled and no plain
+  // observation series. Every series then collapses onto the same
+  // forecast-stripped tooltip key, so no raw series id matches itself.
+  const buildTooltip = (tooltipSortByMetric = false) => {
+    const chartProps = createTestChartProps({
+      formData: {
+        x_axis: 'dt',
+        metrics: seriesIds,
+        groupby: [],
+        richTooltip: true,
+        tooltipSortByMetric,
+      } as Partial<EchartsTimeseriesFormData>,
+      queriesData: [
+        createTestQueryData([
+          {
+            dt: 599616000000,
+            ci__yhat: 1.5,
+            ci__yhat_lower: 0.5,
+            ci__yhat_upper: 2.5,
+          },
+        ]),
+      ],
+    });
+    const tooltipFormatter = (transformProps(chartProps).echartOptions as any)
+      .tooltip.formatter;
+    return tooltipFormatter(
+      seriesIds.map((id, i) => ({
+        seriesId: id,
+        seriesName: id,
+        value: [599616000000, values[i]],
+        data: [599616000000, values[i]],
+        marker,
+      })),
+    );
+  };
+
+  test('renders the collapsed series rather than falling back to "No data"', () => {
+    const html = buildTooltip();
+    expect(html).not.toContain('No data');
+    expect(html).toContain('>ci<');
+    expect(html).toContain('ŷ = 1.5 (0.5, 2.5)');
+  });
+
+  test('renders a single row rather than one per forecast suffix', () => {
+    const html = buildTooltip();
+    expect(html.match(/<tr/g)).toHaveLength(1);
+    expect(html).toContain('>ci<');
+  });
+
+  test('still renders the row when the tooltip is sorted by metric', () => {
+    const html = buildTooltip(true);
+    expect(html).not.toContain('No data');
+    expect(html).toContain('>ci<');
+  });
+});

superset-frontend/plugins/plugin-chart-echarts/test/utils/forecast.test.ts

@@ -23,6 +23,7 @@ import {
 } from '@superset-ui/core';
 import { SeriesOption } from 'echarts';
 import {
+  collapseForecastKeys,
   extractForecastSeriesContext,
   extractForecastValuesFromTooltipParams,
   formatForecastTooltipSeries,
@@ -464,3 +465,35 @@ describe('formatForecastTooltipSeries truncation', () => {
     expect(cell).toBe(`${marker}cpu`);
   });
 });
+
+describe('collapseForecastKeys', () => {
+  test('leaves plain observation series untouched and in order', () => {
+    expect(collapseForecastKeys(['foo', 'bar'])).toEqual(['foo', 'bar']);
+  });
+
+  test('folds a forecast bundle down to a single key', () => {
+    expect(
+      collapseForecastKeys([
+        'foo',
+        'foo__yhat',
+        'foo__yhat_lower',
+        'foo__yhat_upper',
+      ]),
+    ).toEqual(['foo']);
+  });
+
+  test('keeps a key for metrics whose labels are entirely forecast suffixes', () => {
+    // Charts can carry metrics literally labelled `ci__yhat*` with no plain
+    // observation series. Callers match these against forecast-stripped keys,
+    // so an uncollapsed id here would match nothing and drop every row.
+    expect(
+      collapseForecastKeys(['ci__yhat', 'ci__yhat_lower', 'ci__yhat_upper']),
+    ).toEqual(['ci']);
+  });
+
+  test('preserves the incoming order of distinct series', () => {
+    expect(
+      collapseForecastKeys(['b__yhat_lower', 'a__yhat', 'b__yhat']),
+    ).toEqual(['b', 'a']);
+  });
+});