Clarifiied comments and addressed more PR feedbacks.

- pyupgrade: drop forward-reference string annotations in _mcp.py (Python 3.10+ resolves them natively now that MCPTaskOptions is defined before use).

- AGENTS.md: align reconnect description with current behavior. Phase 1 (initial tools/call) does NOT retry on connection loss; raises 'connection lost; task state unknown' instead, so a server that accepted the request but lost the response cannot start the operation twice. Phase 2 (tasks/get / tasks/result) still reconnects once against the same task_id.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

.github/actions/free-runner-disk-space/action.yml

@@ -0,0 +1,64 @@
+name: Free runner disk space
+description: |
+  Reclaims disk space on GitHub-hosted Ubuntu runners by removing
+  pre-installed toolchains we do not use (Android SDK, GHC/Haskell,
+  CodeQL bundle), Docker images, and swap. Also relocates the
+  NuGet package cache to /mnt (which has ~75 GB free vs ~14 GB
+  on /). No-op on non-Linux runners.
+
+runs:
+  using: composite
+  steps:
+    - name: Free disk space (Linux only)
+      if: runner.os == 'Linux'
+      shell: bash
+      run: |
+        set -euo pipefail
+        echo "::group::Disk usage before cleanup"
+        df -h /
+        echo "::endgroup::"
+
+        # Remove pre-installed toolchains we never use on this repo's
+        # dotnet/python jobs. These reclaim ~25-30 GB on ubuntu-latest.
+        sudo rm -rf \
+          /usr/local/lib/android \
+          /usr/share/dotnet/sdk/NuGetFallbackFolder \
+          /opt/ghc \
+          /usr/local/.ghcup \
+          /opt/hostedtoolcache/CodeQL \
+          /opt/hostedtoolcache/PyPy \
+          /opt/hostedtoolcache/Ruby \
+          /opt/hostedtoolcache/go \
+          /usr/local/share/boost \
+          /usr/local/share/powershell \
+          /usr/local/share/chromium \
+          /usr/local/share/vcpkg \
+          /usr/local/lib/heroku \
+          "${AGENT_TOOLSDIRECTORY:-/opt/hostedtoolcache}/PyPy" \
+          "${AGENT_TOOLSDIRECTORY:-/opt/hostedtoolcache}/Ruby" \
+          "${AGENT_TOOLSDIRECTORY:-/opt/hostedtoolcache}/go" || true
+
+        # Drop docker images shipped on the runner; jobs that need
+        # docker pull what they need fresh.
+        if command -v docker >/dev/null 2>&1; then
+          sudo docker image prune --all --force >/dev/null 2>&1 || true
+        fi
+
+        # Disable swap to free its backing file.
+        sudo swapoff -a || true
+        sudo rm -f /mnt/swapfile /swapfile || true
+
+        echo "::group::Disk usage after cleanup"
+        df -h /
+        echo "::endgroup::"
+
+    - name: Relocate NuGet package cache to /mnt (Linux only)
+      if: runner.os == 'Linux'
+      shell: bash
+      run: |
+        set -euo pipefail
+        sudo mkdir -p /mnt/nuget
+        sudo chown -R "$USER":"$USER" /mnt/nuget
+        echo "NUGET_PACKAGES=/mnt/nuget" >> "$GITHUB_ENV"
+        echo "Relocated NuGet package cache to /mnt/nuget"
+        df -h /mnt || true

.github/scripts/pr_limit_moderation.js

@@ -8,6 +8,7 @@ function getPullRequest(context) {
 
   return {
     author: pullRequest.user.login,
+    authorType: pullRequest.user.type,
     labels: pullRequest.labels?.map((label) => label.name).filter(Boolean) ?? [],
     number: pullRequest.number,
   };
@@ -49,6 +50,10 @@ function hasLabel(labels, labelName) {
   return labels.some((label) => label.toLowerCase() === labelName.toLowerCase());
 }
 
+function isDependabotAuthor({ author, authorType }) {
+  return authorType === 'Bot' && author.toLowerCase() === 'dependabot[bot]';
+}
+
 function buildLimitMessage({ author, exemptLabelName, maxOpenPrs, openPrCount }) {
   return [
     `Thank you for your contribution, @${author}.`,
@@ -63,24 +68,37 @@ function buildLimitMessage({ author, exemptLabelName, maxOpenPrs, openPrCount })
 }
 
 async function getOpenPrCount({ github, owner, repo, author, pullRequestNumber }) {
-  const query = `repo:${owner}/${repo} is:pr is:open author:${author}`;
-  const response = await github.rest.search.issuesAndPullRequests({
-    q: query,
+  const openPullRequests = await github.paginate(github.rest.pulls.list, {
+    owner,
+    repo,
+    state: 'open',
     per_page: 100,
   });
 
-  const indexedPrNumbers = response.data.items.map((item) => item.number);
-  const currentPrIsIndexed = indexedPrNumbers.includes(pullRequestNumber);
-  if (currentPrIsIndexed || response.data.total_count >= 100) {
-    return response.data.total_count;
-  }
+  const authorOpenPullRequestNumbers = openPullRequests
+    .filter((pullRequest) => pullRequest.user?.login === author)
+    .map((pullRequest) => pullRequest.number);
+  const currentPrIsOpen = authorOpenPullRequestNumbers.includes(pullRequestNumber);
+  const existingOpenPrCount = currentPrIsOpen
+    ? authorOpenPullRequestNumbers.length - 1
+    : authorOpenPullRequestNumbers.length;
 
-  return response.data.total_count + 1;
+  return existingOpenPrCount + 1;
 }
 
 async function enforcePrLimit({ github, context, core, exemptLabelName, maxOpenPrs, labelName }) {
   const { owner, repo } = context.repo;
-  const { author, labels, number } = getPullRequest(context);
+  const { author, authorType, labels, number } = getPullRequest(context);
+
+  if (isDependabotAuthor({ author, authorType })) {
+    core.info(`Author ${author} is Dependabot; skipping open PR limit enforcement.`);
+    return {
+      author,
+      closed: false,
+      dependabotExempt: true,
+      openPrCount: null,
+    };
+  }
 
   if (hasLabel(labels, exemptLabelName)) {
     core.info(`PR #${number} has the ${exemptLabelName} label; skipping open PR limit enforcement.`);

.github/tests/test_pr_limit_moderation.js

@@ -16,7 +16,7 @@ const { enforcePrLimit } = require('../scripts/pr_limit_moderation.js');
 // Helpers
 // ---------------------------------------------------------------------------
 
-function createContext({ author = 'community-user', labels = [], number = 123 } = {}) {
+function createContext({ author = 'community-user', authorType = 'User', labels = [], number = 123 } = {}) {
   return {
     repo: {
       owner: 'microsoft',
@@ -28,6 +28,7 @@ function createContext({ author = 'community-user', labels = [], number = 123 }
         labels: labels.map((name) => ({ name })),
         user: {
           login: author,
+          type: authorType,
         },
       },
     },
@@ -44,23 +45,20 @@ function createCore() {
   };
 }
 
-function createGithub({ totalCount, itemNumbers, labelExists = true }) {
+function createGithub({
+  itemNumbers,
+  labelExists = true,
+  pullRequests = createPullRequestPage({ numbers: itemNumbers }),
+}) {
   const calls = [];
 
   return {
     calls,
+    async paginate(method, params) {
+      calls.push({ api: 'paginate', method, params });
+      return pullRequests;
+    },
     rest: {
-      search: {
-        async issuesAndPullRequests(params) {
-          calls.push({ api: 'search.issuesAndPullRequests', params });
-          return {
-            data: {
-              total_count: totalCount,
-              items: itemNumbers.map((number) => ({ number })),
-            },
-          };
-        },
-      },
       issues: {
         async getLabel(params) {
           calls.push({ api: 'issues.getLabel', params });
@@ -85,6 +83,10 @@ function createGithub({ totalCount, itemNumbers, labelExists = true }) {
         },
       },
       pulls: {
+        async list(params) {
+          calls.push({ api: 'pulls.list', params });
+          return { data: pullRequests };
+        },
         async update(params) {
           calls.push({ api: 'pulls.update', params });
           return { data: { state: params.state } };
@@ -94,6 +96,15 @@ function createGithub({ totalCount, itemNumbers, labelExists = true }) {
   };
 }
 
+function createPullRequestPage({ author = 'community-user', numbers }) {
+  return numbers.map((number) => ({
+    number,
+    user: {
+      login: author,
+    },
+  }));
+}
+
 
 // ---------------------------------------------------------------------------
 // PR limit enforcement
@@ -102,7 +113,6 @@ function createGithub({ totalCount, itemNumbers, labelExists = true }) {
 describe('PR limit enforcement', () => {
   it('does not close the PR when the author is at the open PR limit', async () => {
     const github = createGithub({
-      totalCount: 10,
       itemNumbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 123],
     });
 
@@ -119,14 +129,13 @@ describe('PR limit enforcement', () => {
     assert.equal(result.openPrCount, 10);
     assert.deepEqual(
       github.calls.map((call) => call.api),
-      ['search.issuesAndPullRequests'],
+      ['paginate'],
     );
   });
 
-  it('counts the new PR when search has not indexed it yet', async () => {
+  it('counts the new PR when the pull list includes it', async () => {
     const github = createGithub({
-      totalCount: 10,
-      itemNumbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
+      itemNumbers: [123, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
     });
 
     const result = await enforcePrLimit({
@@ -143,7 +152,7 @@ describe('PR limit enforcement', () => {
     assert.deepEqual(
       github.calls.map((call) => call.api),
       [
-        'search.issuesAndPullRequests',
+        'paginate',
         'issues.getLabel',
         'issues.addLabels',
         'issues.createComment',
@@ -152,9 +161,31 @@ describe('PR limit enforcement', () => {
     );
   });
 
+  it('counts the current PR on top of existing open PRs', async () => {
+    const github = createGithub({
+      itemNumbers: [123, ...Array.from({ length: 24 }, (_, index) => index + 1)],
+      pullRequests: createPullRequestPage({
+        numbers: [123, ...Array.from({ length: 25 }, (_, index) => index + 1)],
+      }),
+    });
+
+    const result = await enforcePrLimit({
+      github,
+      context: createContext(),
+      core: createCore(),
+      exemptLabelName: 'pr-limit-exempt',
+      maxOpenPrs: 10,
+      labelName: 'too-many-prs',
+    });
+
+    assert.equal(result.closed, true);
+    assert.equal(result.openPrCount, 26);
+    const comment = github.calls.find((call) => call.api === 'issues.createComment').params.body;
+    assert.match(comment, /This PR would put you at 26 open pull requests/);
+  });
+
   it('creates the label when it does not already exist', async () => {
     const github = createGithub({
-      totalCount: 11,
       itemNumbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 123],
       labelExists: false,
     });
@@ -172,7 +203,7 @@ describe('PR limit enforcement', () => {
     assert.deepEqual(
       github.calls.map((call) => call.api),
       [
-        'search.issuesAndPullRequests',
+        'paginate',
         'issues.getLabel',
         'issues.createLabel',
         'issues.addLabels',
@@ -188,7 +219,6 @@ describe('PR limit enforcement', () => {
 
   it('tolerates a 422 race when creating the label', async () => {
     const github = createGithub({
-      totalCount: 11,
       itemNumbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 123],
       labelExists: false,
     });
@@ -212,7 +242,7 @@ describe('PR limit enforcement', () => {
     assert.deepEqual(
       github.calls.map((call) => call.api),
       [
-        'search.issuesAndPullRequests',
+        'paginate',
         'issues.getLabel',
         'issues.createLabel',
         'issues.addLabels',
@@ -224,8 +254,11 @@ describe('PR limit enforcement', () => {
 
   it('uses a diplomatic close message with the configured limit', async () => {
     const github = createGithub({
-      totalCount: 11,
       itemNumbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 123],
+      pullRequests: createPullRequestPage({
+        author: 'octo-contributor',
+        numbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 123],
+      }),
     });
 
     await enforcePrLimit({
@@ -246,7 +279,6 @@ describe('PR limit enforcement', () => {
 
   it('does not close an exempt PR when it is reopened', async () => {
     const github = createGithub({
-      totalCount: 11,
       itemNumbers: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 123],
     });
 
@@ -265,10 +297,33 @@ describe('PR limit enforcement', () => {
     assert.deepEqual(github.calls, []);
   });
 
-  it('does not over-count when the current PR is not on the first search page', async () => {
+  it('does not close Dependabot PRs', async () => {
     const github = createGithub({
-      totalCount: 101,
-      itemNumbers: Array.from({ length: 100 }, (_, index) => index + 1),
+      itemNumbers: [123, ...Array.from({ length: 25 }, (_, index) => index + 1)],
+      pullRequests: createPullRequestPage({
+        author: 'dependabot[bot]',
+        numbers: [123, ...Array.from({ length: 25 }, (_, index) => index + 1)],
+      }),
+    });
+
+    const result = await enforcePrLimit({
+      github,
+      context: createContext({ author: 'dependabot[bot]', authorType: 'Bot' }),
+      core: createCore(),
+      exemptLabelName: 'pr-limit-exempt',
+      maxOpenPrs: 10,
+      labelName: 'too-many-prs',
+    });
+
+    assert.equal(result.closed, false);
+    assert.equal(result.dependabotExempt, true);
+    assert.equal(result.openPrCount, null);
+    assert.deepEqual(github.calls, []);
+  });
+
+  it('counts the current PR when the author has more than one page of open PRs', async () => {
+    const github = createGithub({
+      itemNumbers: [123, ...Array.from({ length: 100 }, (_, index) => index + 1)],
     });
 
     const result = await enforcePrLimit({

.github/workflows/dotnet-build-and-test.yml

@@ -121,6 +121,9 @@ jobs:
               python
               declarative-agents
 
+      - name: Free runner disk space
+        uses: ./.github/actions/free-runner-disk-space
+
       - name: Setup dotnet
         uses: actions/setup-dotnet@c2fa09f4bde5ebb9d1777cf28262a3eb3db3ced7 # v5.2.0
         with:
@@ -191,6 +194,9 @@ jobs:
               python
               declarative-agents
 
+      - name: Free runner disk space
+        uses: ./.github/actions/free-runner-disk-space
+
       # Start Cosmos DB Emulator for all integration tests and only for unit tests when CosmosDB changes happened)
       - name: Start Azure Cosmos DB Emulator
         if: ${{ runner.os == 'Windows' && (needs.paths-filter.outputs.cosmosDbChanges == 'true' || (github.event_name != 'pull_request' && matrix.integration-tests)) }}
@@ -365,6 +371,9 @@ jobs:
             dotnet
             python
 
+      - name: Free runner disk space
+        uses: ./.github/actions/free-runner-disk-space
+
       - name: Setup dotnet
         uses: actions/setup-dotnet@c2fa09f4bde5ebb9d1777cf28262a3eb3db3ced7 # v5.2.0
         with:
@@ -452,6 +461,9 @@ jobs:
             python
             declarative-agents
 
+      - name: Free runner disk space
+        uses: ./.github/actions/free-runner-disk-space
+
       - name: Setup dotnet
         uses: actions/setup-dotnet@c2fa09f4bde5ebb9d1777cf28262a3eb3db3ced7 # v5.2.0
         with:

.github/workflows/python-test-coverage-report.yml

@@ -8,6 +8,7 @@ on:
 
 permissions:
   contents: read
+  actions: read
   pull-requests: write
 
 jobs:
@@ -23,7 +24,7 @@ jobs:
       - name: Download coverage report
         uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
         with:
-          github-token: ${{ secrets.GH_ACTIONS_PR_WRITE }}
+          github-token: ${{ github.token }}
           run-id: ${{ github.event.workflow_run.id }}
           path: ./python
           merge-multiple: true
@@ -38,9 +39,9 @@ jobs:
             echo "PR number file 'pr_number' is missing or empty"
             exit 1
           fi
-          PR_NUMBER=$(head -1 pr_number | tr -dc '0-9')
-          if [ -z "$PR_NUMBER" ]; then
-            echo "PR number file 'pr_number' does not contain a valid PR number"
+          PR_NUMBER=$(cat pr_number)
+          if ! [[ "$PR_NUMBER" =~ ^[0-9]+$ ]]; then
+            echo "::error::PR number file contains invalid content"
             exit 1
           fi
           echo "PR_NUMBER=$PR_NUMBER" >> "$GITHUB_ENV"
@@ -48,7 +49,7 @@ jobs:
         id: coverageComment
         uses: MishaKav/pytest-coverage-comment@26f986d2599c288bb62f623d29c2da98609e9cd4 # v1.6.0
         with:
-          github-token: ${{ secrets.GH_ACTIONS_PR_WRITE }}
+          github-token: ${{ github.token }}
           issue-number: ${{ env.PR_NUMBER }}
           pytest-xml-coverage-path: python/python-coverage.xml
           title: "Python Test Coverage Report"

.gitignore

@@ -248,3 +248,4 @@ dotnet/filtered-*.slnx
 .omx/
 
 **/issues/
+.test_*

SUPPORT.md

@@ -1,17 +1,17 @@
-# Support
-
-## How to file issues and get help  
-
-This project uses GitHub Issues to track bugs and feature requests. Please search the existing 
-issues before filing new issues to avoid duplicates.  For new issues, file your bug or 
-feature request as a new Issue.
-
-For help and questions about using this project, please create a GitHub issue.
-
-AI Support team will support Microsoft Agent Framework issues for customers under a **Unified support agreement when the issue arises from usage of Azure AI services** (Foundry Models, Foundry Agents etc.) in conjunction with the SDK. Conversely, if customer has any other / non unified support agreement and/or Agent Framework SDK is used in a way **not involving an Azure service**, it is treated as a purely open-source tool – Microsoft’s support organization will not handle it, and users should use GitHub or forums for assistance
-
-For Copilot Studio SDK implementation issues, customers should use GitHub Issues for assistance, as outlined above. Conversely, for prerequisites managed within the Copilot Studio portal, customers can rely on the standard Microsoft Copilot Studio support channels.
-
-## Microsoft Support Policy  
-
-Support for this **PROJECT or PRODUCT** is limited to the resources listed above.
+# Support
+
+## How to file issues and get help  
+
+This project uses GitHub Issues to track bugs and feature requests. Please search the existing 
+issues before filing new issues to avoid duplicates.  For new issues, file your bug or 
+feature request as a new Issue.
+
+For help and questions about using this project, please create a GitHub issue.
+
+AI Support team will support Microsoft Agent Framework issues for customers under a **Unified support agreement when the issue arises from usage of Azure AI services** (Foundry Models, Foundry Agents etc.) in conjunction with the SDK. Conversely, if customer has any other / non unified support agreement and/or Agent Framework SDK is used in a way **not involving an Azure service**, it is treated as a purely open-source tool – Microsoft’s support organization will not handle it, and users should use GitHub or forums for assistance
+
+For Copilot Studio SDK implementation issues, customers should use GitHub Issues for assistance, as outlined above. Conversely, for prerequisites managed within the Copilot Studio portal, customers can rely on the standard Microsoft Copilot Studio support channels.
+
+## Microsoft Support Policy  
+
+Support for this **PROJECT or PRODUCT** is limited to the resources listed above.

dotnet/agent-framework-dotnet.slnx

@@ -344,6 +344,9 @@
   <Folder Name="/Samples/04-hosting/FoundryHostedAgents/responses/Hosted-Toolbox/">
     <Project Path="samples/04-hosting/FoundryHostedAgents/responses/Hosted-Toolbox/HostedToolbox.csproj" />
   </Folder>
+  <Folder Name="/Samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/">
+    <Project Path="samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/HostedToolboxMcpSkills.csproj" />
+  </Folder>
   <Folder Name="/Samples/04-hosting/FoundryHostedAgents/responses/Hosted-AzureSearchRag/">
     <Project Path="samples/04-hosting/FoundryHostedAgents/responses/Hosted-AzureSearchRag/HostedAzureSearchRag.csproj" />
   </Folder>

dotnet/nuget/nuget-package.props

@@ -1,14 +1,14 @@
 <Project>
   <PropertyGroup>
     <!-- Central version prefix - applies to all nuget packages. -->
-    <VersionPrefix>1.8.0</VersionPrefix>
+    <VersionPrefix>1.9.0</VersionPrefix>
     <RCNumber>1</RCNumber>
-    <DateSuffix>260528</DateSuffix>
+    <DateSuffix>260603</DateSuffix>
     <PackageVersion Condition="'$(IsReleaseCandidate)' == 'true'">$(VersionPrefix)-rc$(RCNumber)</PackageVersion>
     <PackageVersion Condition="'$(IsReleaseCandidate)' != 'true' AND '$(VersionSuffix)' != ''">$(VersionPrefix)-$(VersionSuffix).$(DateSuffix).1</PackageVersion>
     <PackageVersion Condition="'$(IsReleaseCandidate)' != 'true' AND '$(VersionSuffix)' == ''">$(VersionPrefix)-preview.$(DateSuffix).1</PackageVersion>
     <PackageVersion Condition="'$(IsReleased)' == 'true'">$(VersionPrefix)</PackageVersion>
-    <GitTag>1.8.0</GitTag>
+    <GitTag>1.9.0</GitTag>
 
     <Configurations>Debug;Release;Publish</Configurations>
     <IsPackable>true</IsPackable>

dotnet/samples/02-agents/AGUI/Step01_GettingStarted/Server/Program.cs

@@ -10,6 +10,11 @@ WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
 builder.Services.AddHttpClient().AddLogging();
 builder.Services.AddAGUI();
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 string endpoint = builder.Configuration["AZURE_OPENAI_ENDPOINT"]

dotnet/samples/02-agents/AGUI/Step01_GettingStarted/Server/Server.csproj

@@ -14,6 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>
 

dotnet/samples/02-agents/AGUI/Step02_BackendTools/Server/Program.cs

@@ -16,6 +16,11 @@ builder.Services.ConfigureHttpJsonOptions(options =>
     options.SerializerOptions.TypeInfoResolverChain.Add(SampleJsonSerializerContext.Default));
 builder.Services.AddAGUI();
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 string endpoint = builder.Configuration["AZURE_OPENAI_ENDPOINT"]

dotnet/samples/02-agents/AGUI/Step02_BackendTools/Server/Server.csproj

@@ -14,6 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>
 

dotnet/samples/02-agents/AGUI/Step03_FrontendTools/Server/Program.cs

@@ -10,6 +10,11 @@ WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
 builder.Services.AddHttpClient().AddLogging();
 builder.Services.AddAGUI();
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 string endpoint = builder.Configuration["AZURE_OPENAI_ENDPOINT"]

dotnet/samples/02-agents/AGUI/Step03_FrontendTools/Server/Server.csproj

@@ -14,6 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>
 

dotnet/samples/02-agents/AGUI/Step04_HumanInLoop/Server/Program.cs

@@ -27,6 +27,11 @@ builder.Services.ConfigureHttpJsonOptions(options =>
     options.SerializerOptions.TypeInfoResolverChain.Add(ApprovalJsonContext.Default));
 builder.Services.AddAGUI();
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 app.UseHttpLogging();

dotnet/samples/02-agents/AGUI/Step04_HumanInLoop/Server/Server.csproj

@@ -14,6 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>
 

dotnet/samples/02-agents/AGUI/Step05_StateManagement/Server/Program.cs

@@ -17,6 +17,11 @@ builder.Services.AddAGUI();
 // Configure to listen on port 8888
 builder.WebHost.UseUrls("http://localhost:8888");
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 string endpoint = builder.Configuration["AZURE_OPENAI_ENDPOINT"]

dotnet/samples/02-agents/AGUI/Step05_StateManagement/Server/Server.csproj

@@ -14,6 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>
 

dotnet/samples/03-workflows/Observability/WorkflowAsAnAgent/WorkflowHelper.cs

@@ -50,12 +50,16 @@ internal static partial class WorkflowHelper
     /// <summary>
     /// Executor that starts the concurrent processing by sending messages to the agents.
     /// </summary>
-    private sealed partial class ConcurrentStartExecutor() : Executor("ConcurrentStartExecutor")
+    [SendsMessage(typeof(List<ChatMessage>))]
+    [SendsMessage(typeof(TurnToken))]
+    private sealed partial class ConcurrentStartExecutor()
+        : Executor("ConcurrentStartExecutor", declareCrossRunShareable: true), IResettableExecutor
     {
         [MessageHandler]
-        internal ValueTask RouteMessages(List<ChatMessage> messages, IWorkflowContext context, CancellationToken cancellationToken)
+        internal ValueTask RouteMessages(IEnumerable<ChatMessage> messages, IWorkflowContext context, CancellationToken cancellationToken)
         {
-            return context.SendMessageAsync(messages, cancellationToken: cancellationToken);
+            List<ChatMessage> payload = messages as List<ChatMessage> ?? messages.ToList();
+            return context.SendMessageAsync(payload, cancellationToken: cancellationToken);
         }
 
         [MessageHandler]
@@ -63,13 +67,16 @@ internal static partial class WorkflowHelper
         {
             return context.SendMessageAsync(token, cancellationToken: cancellationToken);
         }
+
+        public ValueTask ResetAsync() => default;
     }
 
     /// <summary>
     /// Executor that aggregates the results from the concurrent agents.
     /// </summary>
-    [YieldsOutput(typeof(List<ChatMessage>))]
-    private sealed partial class ConcurrentAggregationExecutor() : Executor<List<ChatMessage>>("ConcurrentAggregationExecutor")
+    [YieldsOutput(typeof(string))]
+    private sealed partial class ConcurrentAggregationExecutor() :
+        Executor<List<ChatMessage>>("ConcurrentAggregationExecutor"), IResettableExecutor
     {
         private readonly List<ChatMessage> _messages = [];
 
@@ -90,5 +97,11 @@ internal static partial class WorkflowHelper
                 await context.YieldOutputAsync(formattedMessages, cancellationToken);
             }
         }
+
+        public ValueTask ResetAsync()
+        {
+            this._messages.Clear();
+            return default;
+        }
     }
 }

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/01_SingleAgent/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/02_AgentOrchestration_Chaining/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/03_AgentOrchestration_Concurrency/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/04_AgentOrchestration_Conditionals/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/05_AgentOrchestration_HITL/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/06_LongRunningTools/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/07_AgentAsMcpTool/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableAgents/AzureFunctions/08_ReliableStreaming/local.settings.json

@@ -1,12 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>",
-    "REDIS_CONNECTION_STRING": "localhost:6379",
-    "REDIS_STREAM_TTL_MINUTES": "10"
-  }
-}

dotnet/samples/04-hosting/DurableWorkflows/AzureFunctions/01_SequentialWorkflow/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableWorkflows/AzureFunctions/02_ConcurrentWorkflow/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableWorkflows/AzureFunctions/03_WorkflowHITL/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/DurableWorkflows/AzureFunctions/04_WorkflowMcpTool/local.settings.json

@@ -1,8 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None"
-  }
-}

dotnet/samples/04-hosting/DurableWorkflows/AzureFunctions/05_WorkflowAndAgents/local.settings.json

@@ -1,10 +0,0 @@
-{
-  "IsEncrypted": false,
-  "Values": {
-    "FUNCTIONS_WORKER_RUNTIME": "dotnet-isolated",
-    "AzureWebJobsStorage": "UseDevelopmentStorage=true",
-    "DURABLE_TASK_SCHEDULER_CONNECTION_STRING": "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None",
-    "AZURE_OPENAI_ENDPOINT": "<AZURE_OPENAI_ENDPOINT>",
-    "AZURE_OPENAI_DEPLOYMENT_NAME": "<AZURE_OPENAI_DEPLOYMENT_NAME>"
-  }
-}

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/.env.example

@@ -0,0 +1,6 @@
+AZURE_AI_PROJECT_ENDPOINT=<your-azure-ai-project-endpoint>
+ASPNETCORE_URLS=http://+:8088
+ASPNETCORE_ENVIRONMENT=Development
+AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-5
+FOUNDRY_TOOLBOX_NAME=<your-toolbox-name>
+AZURE_BEARER_TOKEN=DefaultAzureCredential

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/Dockerfile

@@ -0,0 +1,26 @@
+# Dockerfile for end-users consuming the Agent Framework via NuGet packages.
+#
+# This Dockerfile performs a full `dotnet restore` and `dotnet publish` inside the container,
+# which only succeeds when the project references its dependencies via PackageReference (see the
+# commented-out section in HostedToolboxMcpSkills.csproj). Contributors building from the
+# agent-framework repository source must use Dockerfile.contributor instead because
+# ProjectReference dependencies live outside this folder and cannot be restored from inside
+# this build context.
+#
+# Use the official .NET 10.0 ASP.NET runtime as a parent image
+FROM mcr.microsoft.com/dotnet/aspnet:10.0 AS base
+WORKDIR /app
+
+FROM mcr.microsoft.com/dotnet/sdk:10.0 AS build
+WORKDIR /src
+COPY . .
+RUN dotnet restore
+RUN dotnet publish -c Release -o /app/publish
+
+# Final stage
+FROM base AS final
+WORKDIR /app
+COPY --from=build /app/publish .
+EXPOSE 8088
+ENV ASPNETCORE_URLS=http://+:8088
+ENTRYPOINT ["dotnet", "HostedToolboxMcpSkills.dll"]

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/Dockerfile.contributor

@@ -0,0 +1,18 @@
+# Dockerfile for contributors building from the agent-framework repository source.
+#
+# This project uses ProjectReference to the local source, which means a standard
+# multi-stage Docker build cannot resolve dependencies outside this folder.
+# Pre-publish the app targeting the container runtime and copy the output:
+#
+#   dotnet publish -c Debug -f net10.0 -r linux-musl-x64 --self-contained false -o out
+#   docker build -f Dockerfile.contributor -t hosted-toolbox-mcp-skills .
+#   docker run --rm -p 8088:8088 -e AGENT_NAME=hosted-toolbox-mcp-skills -e AZURE_BEARER_TOKEN=$AZURE_BEARER_TOKEN --env-file .env hosted-toolbox-mcp-skills
+#
+# For end-users consuming the NuGet package (not ProjectReference), use the standard
+# Dockerfile which performs a full dotnet restore + publish inside the container.
+FROM mcr.microsoft.com/dotnet/aspnet:10.0-alpine AS final
+WORKDIR /app
+COPY out/ .
+EXPOSE 8088
+ENV ASPNETCORE_URLS=http://+:8088
+ENTRYPOINT ["dotnet", "HostedToolboxMcpSkills.dll"]

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/HostedToolboxMcpSkills.csproj

@@ -0,0 +1,36 @@
+<Project Sdk="Microsoft.NET.Sdk.Web">
+
+  <PropertyGroup>
+    <TargetFrameworks>net10.0</TargetFrameworks>
+    <Nullable>enable</Nullable>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <CentralPackageTransitivePinningEnabled>false</CentralPackageTransitivePinningEnabled>
+    <RootNamespace>HostedToolboxMcpSkills</RootNamespace>
+    <AssemblyName>HostedToolboxMcpSkills</AssemblyName>
+    <NoWarn>$(NoWarn);</NoWarn>
+  </PropertyGroup>
+
+  <ItemGroup>
+    <PackageReference Include="Azure.AI.Projects" />
+    <PackageReference Include="Azure.Identity" />
+    <PackageReference Include="ModelContextProtocol" VersionOverride="1.2.0" />
+    <PackageReference Include="DotNetEnv" />
+  </ItemGroup>
+
+  <!-- For contributors: uses ProjectReference to build against local source -->
+  <ItemGroup>
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Foundry\Microsoft.Agents.AI.Foundry.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Foundry.Hosting\Microsoft.Agents.AI.Foundry.Hosting.csproj" />
+    <ProjectReference Include="..\..\..\..\..\src\Microsoft.Agents.AI.Mcp\Microsoft.Agents.AI.Mcp.csproj" />
+    <ProjectReference Include="..\Hosted_Shared_Contributor_Setup\Hosted_Shared_Contributor_Setup.csproj" />
+  </ItemGroup>
+
+  <!-- For end-users: uncomment the PackageReference below and remove the ProjectReference above
+  <ItemGroup>
+    <PackageReference Include="Microsoft.Agents.AI.Foundry" Version="1.6.1-preview.260514.1" />
+    <PackageReference Include="Microsoft.Agents.AI.Foundry.Hosting" Version="1.6.1-preview.260514.1" />
+    <PackageReference Include="Microsoft.Agents.AI.Mcp" Version="1.6.1-preview.260514.1" />
+  </ItemGroup>
+  -->
+
+</Project>

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/Program.cs

@@ -0,0 +1,109 @@
+// Copyright (c) Microsoft. All rights reserved.
+
+// Hosted Toolbox MCP Skills Agent
+//
+// Demonstrates how to host an agent that discovers MCP-based skills from a
+// Foundry Toolbox MCP endpoint and injects them as AIContextProviders using
+// AgentSkillsProviderBuilder.UseMcpSkills().
+//
+// Required environment variables:
+//   AZURE_AI_PROJECT_ENDPOINT         - Azure AI Foundry project endpoint
+//   FOUNDRY_TOOLBOX_NAME              - Name of the Foundry Toolbox to connect to
+//   AZURE_AI_MODEL_DEPLOYMENT_NAME    - Model deployment name (default: gpt-5)
+
+using System.Net.Http.Headers;
+using Azure.AI.Projects;
+using Azure.Core;
+using Azure.Identity;
+using DotNetEnv;
+using Hosted_Shared_Contributor_Setup;
+using Microsoft.Agents.AI;
+using Microsoft.Agents.AI.Foundry.Hosting;
+using ModelContextProtocol.Client;
+
+// Load .env file if present (for local development)
+Env.TraversePath().Load();
+
+var projectEndpoint = Environment.GetEnvironmentVariable("AZURE_AI_PROJECT_ENDPOINT")
+    ?? throw new InvalidOperationException("AZURE_AI_PROJECT_ENDPOINT is not set.");
+var deployment = Environment.GetEnvironmentVariable("AZURE_AI_MODEL_DEPLOYMENT_NAME") ?? "gpt-5";
+var toolboxName = Environment.GetEnvironmentVariable("FOUNDRY_TOOLBOX_NAME")
+    ?? throw new InvalidOperationException("FOUNDRY_TOOLBOX_NAME is not set.");
+
+// Build the Toolbox MCP URL from the project endpoint and toolbox name.
+var toolboxMcpServerUrl = $"{projectEndpoint.TrimEnd('/')}/toolboxes/{toolboxName}/mcp?api-version=v1";
+
+// Use a chained credential: try a temporary dev token first (for local Docker debugging),
+// then fall back to DefaultAzureCredential (for local dev via dotnet run / managed identity in production).
+TokenCredential credential = new ChainedTokenCredential(
+    new DevTemporaryTokenCredential(),
+    new DefaultAzureCredential());
+
+// ── Connect to the Foundry Toolbox MCP endpoint ─────────────────────────────
+// Create an HttpClient that attaches a fresh Foundry bearer token to every request.
+using var httpClient = new HttpClient(new BearerTokenHandler(credential, "https://ai.azure.com/.default") { CheckCertificateRevocationList = true });
+
+Console.WriteLine($"Connecting to Foundry Toolbox '{toolboxName}' MCP server...");
+
+await using var mcpClient = await McpClient.CreateAsync(
+    new HttpClientTransport(
+        new HttpClientTransportOptions
+        {
+            Endpoint = new Uri(toolboxMcpServerUrl),
+            Name = toolboxName,
+            TransportMode = HttpTransportMode.StreamableHttp,
+            AdditionalHeaders = new Dictionary<string, string>
+            {
+                ["Foundry-Features"] = "Toolboxes=V1Preview",
+            },
+        },
+        httpClient));
+
+// ── Configure MCP-based skills provider ──────────────────────────────────────
+var skillsProvider = new AgentSkillsProviderBuilder()
+    .UseMcpSkills(mcpClient)
+    .Build();
+
+// ── Create the agent ─────────────────────────────────────────────────────────
+AIAgent agent = new AIProjectClient(new Uri(projectEndpoint), credential)
+    .AsAIAgent(new ChatClientAgentOptions
+    {
+        Name = Environment.GetEnvironmentVariable("AGENT_NAME") ?? "hosted-toolbox-mcp-skills",
+        Description = "Hosted agent with MCP skills discovered from a Foundry Toolbox",
+        ChatOptions = new()
+        {
+            ModelId = deployment,
+            Instructions = "You are a helpful assistant.",
+        },
+        AIContextProviders = [skillsProvider],
+    });
+
+// ── Build the host ───────────────────────────────────────────────────────────
+var builder = WebApplication.CreateBuilder(args);
+builder.Services.AddFoundryResponses(agent);
+builder.Services.AddDevTemporaryLocalContributorSetup(); // Local Docker debugging only - must not be used in production.
+
+var app = builder.Build();
+app.MapFoundryResponses();
+
+// Contributor-only: in Development, also map the per-agent OpenAI route shape that live Foundry uses
+// so a local REPL client can target this server via AIProjectClient.AsAIAgent(Uri agentEndpoint).
+// Do not use this in production. Hosted Foundry agents only support the agent-endpoint path.
+app.MapDevTemporaryLocalAgentEndpoint();
+
+app.Run();
+
+// ---------------------------------------------------------------------------
+// HttpClientHandler: attaches a fresh Foundry bearer token to every request
+// ---------------------------------------------------------------------------
+internal sealed class BearerTokenHandler(TokenCredential credential, string scope) : HttpClientHandler
+{
+    private readonly TokenRequestContext _tokenContext = new([scope]);
+
+    protected override async Task<HttpResponseMessage> SendAsync(HttpRequestMessage request, CancellationToken cancellationToken)
+    {
+        AccessToken token = await credential.GetTokenAsync(this._tokenContext, cancellationToken).ConfigureAwait(false);
+        request.Headers.Authorization = new AuthenticationHeaderValue("Bearer", token.Token);
+        return await base.SendAsync(request, cancellationToken).ConfigureAwait(false);
+    }
+}

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/README.md

@@ -0,0 +1,103 @@
+# Hosted-ToolboxMcpSkills
+
+A hosted agent that discovers **MCP-based skills from a Foundry Toolbox** and makes them available to the agent using `AgentSkillsProviderBuilder.UseMcpSkills(mcpClient)`.
+
+The `AgentSkillsProvider` is attached to the agent as a context provider and implements the [Agent Skills](https://agentskills.io/) progressive-disclosure pattern. When the agent is prompted, it discovers available skills in the Foundry Toolbox via the provider:
+
+1. **Advertise** - skill names and descriptions are injected into the system prompt so the agent knows what is available.
+2. **Load** - when the agent decides a skill is relevant, it retrieves the full skill body with detailed instructions via the provider.
+3. **Read resources** - if a skill includes supplementary content (reference documents, assets), the agent reads them on demand via the provider.
+
+This way the full skill body and resources are only loaded when the agent actually needs them, reducing token usage.
+
+## Prerequisites
+
+- [.NET 10 SDK](https://dotnet.microsoft.com/download/dotnet/10.0)
+- An Azure AI Foundry project with a deployed model (e.g., `gpt-5`)
+- A Foundry Toolbox already configured with skills provisioned
+- Azure CLI logged in (`az login`)
+
+## Configuration
+
+Copy the template and fill in your values:
+
+```bash
+cp .env.example .env
+```
+
+Edit `.env` and set your Azure AI Foundry project endpoint and toolbox name:
+
+```env
+AZURE_AI_PROJECT_ENDPOINT=https://<your-account>.services.ai.azure.com/api/projects/<your-project>
+ASPNETCORE_URLS=http://+:8088
+ASPNETCORE_ENVIRONMENT=Development
+AZURE_AI_MODEL_DEPLOYMENT_NAME=gpt-5
+FOUNDRY_TOOLBOX_NAME=my-toolbox
+```
+
+> **Note:** `.env` is gitignored. The `.env.example` template is checked in as a reference.
+
+## Running directly (contributors)
+
+This project uses `ProjectReference` to build against the local Agent Framework source.
+
+```bash
+cd dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills
+dotnet run
+```
+
+The agent will start on `http://localhost:8088`.
+
+### Test it
+
+Using the Azure Developer CLI:
+
+```bash
+azd ai agent invoke --local "What skills do you have available?"
+```
+
+## Running with Docker
+
+Since this project uses `ProjectReference`, use `Dockerfile.contributor` which takes a pre-published output.
+
+### 1. Publish for the container runtime (Linux Alpine)
+
+```bash
+dotnet publish -c Debug -f net10.0 -r linux-musl-x64 --self-contained false -o out
+```
+
+### 2. Build the Docker image
+
+```bash
+docker build -f Dockerfile.contributor -t hosted-toolbox-mcp-skills .
+```
+
+### 3. Run the container
+
+Generate a bearer token on your host and pass it to the container:
+
+```bash
+# Generate token (expires in ~1 hour)
+export AZURE_BEARER_TOKEN=$(az account get-access-token --resource https://ai.azure.com --query accessToken -o tsv)
+
+# Run with token
+docker run --rm -p 8088:8088 \
+  -e AGENT_NAME=hosted-toolbox-mcp-skills \
+  -e AZURE_BEARER_TOKEN=$AZURE_BEARER_TOKEN \
+  --env-file .env \
+  hosted-toolbox-mcp-skills
+```
+
+> **Note:** `AGENT_NAME` is passed via `-e` to simulate the platform injection. `AZURE_BEARER_TOKEN` provides Azure credentials to the container (tokens expire after ~1 hour). The `.env` file provides the remaining configuration.
+
+### 4. Test it
+
+Using the Azure Developer CLI:
+
+```bash
+azd ai agent invoke --local "What skills do you have available?"
+```
+
+## NuGet package users
+
+If you are consuming the Agent Framework as a NuGet package (not building from source), use the standard `Dockerfile` instead of `Dockerfile.contributor`. See the commented section in `HostedToolboxMcpSkills.csproj` for the `PackageReference` alternative.

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/agent.manifest.yaml

@@ -0,0 +1,43 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/microsoft/AgentSchema/refs/heads/main/schemas/v1.0/AgentManifest.yaml
+name: hosted-toolbox-mcp-skills
+displayName: "Hosted Toolbox MCP Skills Agent"
+
+description: >
+  A hosted agent that discovers MCP-based skills from a Foundry Toolbox
+  and makes them available to the agent via the agent skills provider.
+
+metadata:
+  tags:
+    - AI Agent Hosting
+    - Azure AI AgentServer
+    - Responses Protocol
+    - Agent Framework
+    - MCP
+    - Model Context Protocol
+    - Agent Skills
+    - Foundry Toolbox
+    - Foundry Toolbox Skills
+
+template:
+  name: hosted-toolbox-mcp-skills
+  kind: hosted
+  protocols:
+    - protocol: responses
+      version: 1.0.0
+  resources:
+    cpu: "0.25"
+    memory: 0.5Gi
+  environment_variables:
+    - name: AZURE_AI_MODEL_DEPLOYMENT_NAME
+      value: "{{AZURE_AI_MODEL_DEPLOYMENT_NAME}}"
+    - name: FOUNDRY_TOOLBOX_NAME
+      value: "{{FOUNDRY_TOOLBOX_NAME}}"
+parameters:
+  properties:
+    - name: FOUNDRY_TOOLBOX_NAME
+      secret: false
+      description: Name of the Foundry Toolbox to connect to for MCP skill discovery
+resources:
+  - kind: model
+    id: gpt-5
+    name: AZURE_AI_MODEL_DEPLOYMENT_NAME

dotnet/samples/04-hosting/FoundryHostedAgents/responses/Hosted-ToolboxMcpSkills/agent.yaml

@@ -0,0 +1,14 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/microsoft/AgentSchema/refs/heads/main/schemas/v1.0/ContainerAgent.yaml
+kind: hosted
+name: hosted-toolbox-mcp-skills
+protocols:
+  - protocol: responses
+    version: 1.0.0
+resources:
+  cpu: "0.25"
+  memory: 0.5Gi
+environment_variables:
+  - name: AZURE_AI_MODEL_DEPLOYMENT_NAME
+    value: ${AZURE_AI_MODEL_DEPLOYMENT_NAME}
+  - name: FOUNDRY_TOOLBOX_NAME
+    value: ${FOUNDRY_TOOLBOX_NAME}

dotnet/samples/05-end-to-end/AGUIClientServer/AGUIDojoServer/AGUIDojoServer.csproj

@@ -15,6 +15,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>
 

dotnet/samples/05-end-to-end/AGUIClientServer/AGUIDojoServer/Program.cs

@@ -19,6 +19,11 @@ builder.Services.AddHttpClient().AddLogging();
 builder.Services.ConfigureHttpJsonOptions(options => options.SerializerOptions.TypeInfoResolverChain.Add(AGUIDojoServerSerializerContext.Default));
 builder.Services.AddAGUI();
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 app.UseHttpLogging();

dotnet/samples/05-end-to-end/AGUIClientServer/AGUIServer/Program.cs

@@ -49,8 +49,9 @@ var agent = new AzureOpenAIClient(
                 AGUIServerSerializerContext.Default.Options)
         ]);
 
-// When running in production, make sure to use an SessionIsolationKeyProvider, e.g. ClaimsIdentity-based
-// if using Claims-based Identity for Authentication/Authorization
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
 // builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
 
 // Register the agent with the host and configure it to use an in-memory session store

dotnet/samples/05-end-to-end/AGUIWebChat/Server/AGUIWebChatServer.csproj

@@ -14,6 +14,7 @@
 
   <ItemGroup>
     <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore\Microsoft.Agents.AI.Hosting.AGUI.AspNetCore.csproj" />
+    <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.Hosting.AspNetCore\Microsoft.Agents.AI.Hosting.AspNetCore.csproj" />
     <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.AGUI\Microsoft.Agents.AI.AGUI.csproj" />
     <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
   </ItemGroup>

dotnet/samples/05-end-to-end/AGUIWebChat/Server/Program.cs

@@ -12,6 +12,11 @@ WebApplicationBuilder builder = WebApplication.CreateBuilder(args);
 builder.Services.AddHttpClient().AddLogging();
 builder.Services.AddAGUI();
 
+// WARNING: When adding session persistence (e.g., WithInMemorySessionStore), or running in production,
+// make sure to also register a SessionIsolationKeyProvider to scope sessions by principal in multi-user
+// deployments, e.g.:
+// builder.Services.UseClaimsBasedSessionIsolation(new() { ClaimType = ClaimTypes.NameIdentifier });
+
 WebApplication app = builder.Build();
 
 string endpoint = builder.Configuration["AZURE_OPENAI_ENDPOINT"] ?? throw new InvalidOperationException("AZURE_OPENAI_ENDPOINT is not set.");

dotnet/src/Microsoft.Agents.AI.Foundry.Hosting/OutputConverter.cs

@@ -281,14 +281,19 @@ internal static class OutputConverter
 
                         var outputText = EncodeFunctionResultAsJsonStringPayload(functionResult.Result);
 
-                        var itemId = GenerateItemId("fc");
-                        var outputItem = new OutputItemFunctionToolCallOutput(
+                        // Use the SDK's convenience method so the OutputItemFunctionToolCallOutput
+                        // is constructed with a populated Id. The public OutputItemFunctionToolCallOutput
+                        // ctor only sets CallId/Output (Id is read-only), and AddOutputItem<T>+EmitAdded
+                        // does not auto-stamp Id — only ResponseId/AgentReference. Without this, the
+                        // serialized item arrives at the Foundry storage layer with id=null and is
+                        // rejected with "ID cannot be null or empty (Parameter 'id')".
+                        foreach (var evt in stream.OutputItemFunctionCallOutput(
                             functionResult.CallId,
-                            BinaryData.FromString(outputText));
+                            BinaryData.FromString(outputText)))
+                        {
+                            yield return evt;
+                        }
 
-                        var outputBuilder = stream.AddOutputItem<OutputItemFunctionToolCallOutput>(itemId);
-                        yield return outputBuilder.EmitAdded(outputItem);
-                        yield return outputBuilder.EmitDone(outputItem);
                         break;
                     }
 

dotnet/src/Microsoft.Agents.AI.Foundry/Microsoft.Agents.AI.Foundry.csproj

@@ -24,11 +24,13 @@
 
   <ItemGroup>
     <PackageReference Include="Azure.AI.Projects" />
+    <PackageReference Include="Azure.Core" />
     <PackageReference Include="Microsoft.Extensions.AI" />
     <PackageReference Include="Microsoft.Extensions.AI.Abstractions" />
     <PackageReference Include="Microsoft.Extensions.AI.OpenAI" />
     <PackageReference Include="Microsoft.Extensions.Compliance.Abstractions" />
     <PackageReference Include="OpenAI" />
+    <PackageReference Include="System.ClientModel" />
   </ItemGroup>
 
   <!-- Evaluation support requires net8.0+ (MEAI.Evaluation does not support legacy TFMs) -->

dotnet/src/Microsoft.Agents.AI.Harness/ChatClientHarnessExtensions.cs

@@ -1,7 +1,9 @@
 // Copyright (c) Microsoft. All rights reserved.
 
+using System;
 using System.Diagnostics.CodeAnalysis;
 using Microsoft.Agents.AI;
+using Microsoft.Extensions.Logging;
 using Microsoft.Shared.DiagnosticIds;
 
 namespace Microsoft.Extensions.AI;
@@ -32,11 +34,19 @@ public static class ChatClientHarnessExtensions
     /// additional context providers, and chat history provider.
     /// When <see langword="null"/>, the agent uses built-in default settings.
     /// </param>
+    /// <param name="loggerFactory">
+    /// Optional logger factory for creating loggers used by the agent and its components.
+    /// </param>
+    /// <param name="services">
+    /// Optional service provider for resolving dependencies required by AI functions and other agent components.
+    /// </param>
     /// <returns>A new <see cref="HarnessAgent"/> instance.</returns>
     public static HarnessAgent AsHarnessAgent(
         this IChatClient chatClient,
         int maxContextWindowTokens,
         int maxOutputTokens,
-        HarnessAgentOptions? options = null) =>
-        new(chatClient, maxContextWindowTokens, maxOutputTokens, options);
+        HarnessAgentOptions? options = null,
+        ILoggerFactory? loggerFactory = null,
+        IServiceProvider? services = null) =>
+        new(chatClient, maxContextWindowTokens, maxOutputTokens, options, loggerFactory, services);
 }

dotnet/src/Microsoft.Agents.AI.Harness/HarnessAgent.cs

@@ -10,6 +10,7 @@ using Microsoft.Agents.AI.Compaction;
 using Microsoft.Agents.AI.Tools.Shell;
 #endif
 using Microsoft.Extensions.AI;
+using Microsoft.Extensions.Logging;
 using Microsoft.Shared.DiagnosticIds;
 using Microsoft.Shared.Diagnostics;
 
@@ -105,6 +106,12 @@ public sealed class HarnessAgent : DelegatingAIAgent
     /// additional context providers, and chat history provider.
     /// When <see langword="null"/>, the agent uses built-in default settings.
     /// </param>
+    /// <param name="loggerFactory">
+    /// Optional logger factory for creating loggers used by the agent and its components.
+    /// </param>
+    /// <param name="services">
+    /// Optional service provider for resolving dependencies required by AI functions and other agent components.
+    /// </param>
     /// <exception cref="ArgumentNullException">
     /// <paramref name="chatClient"/> is <see langword="null"/>.
     /// </exception>
@@ -112,18 +119,20 @@ public sealed class HarnessAgent : DelegatingAIAgent
     /// <paramref name="maxContextWindowTokens"/> is not positive, or
     /// <paramref name="maxOutputTokens"/> is negative or greater than or equal to <paramref name="maxContextWindowTokens"/>.
     /// </exception>
-    public HarnessAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options = null)
+    public HarnessAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options = null, ILoggerFactory? loggerFactory = null, IServiceProvider? services = null)
         : base(BuildAgent(
             Throw.IfNull(chatClient),
             maxContextWindowTokens,
             maxOutputTokens,
-            options))
+            options,
+            loggerFactory,
+            services))
     {
     }
 
-    private static AIAgent BuildAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options)
+    private static AIAgent BuildAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options, ILoggerFactory? loggerFactory, IServiceProvider? services)
     {
-        ChatClientAgent innerAgent = BuildInnerAgent(chatClient, maxContextWindowTokens, maxOutputTokens, options);
+        ChatClientAgent innerAgent = BuildInnerAgent(chatClient, maxContextWindowTokens, maxOutputTokens, options, loggerFactory, services);
 
         AIAgentBuilder builder = innerAgent.AsBuilder();
 
@@ -137,10 +146,10 @@ public sealed class HarnessAgent : DelegatingAIAgent
             builder.UseOpenTelemetry(sourceName: options?.OpenTelemetrySourceName);
         }
 
-        return builder.Build();
+        return builder.Build(services);
     }
 
-    private static ChatClientAgent BuildInnerAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options)
+    private static ChatClientAgent BuildInnerAgent(IChatClient chatClient, int maxContextWindowTokens, int maxOutputTokens, HarnessAgentOptions? options, ILoggerFactory? loggerFactory, IServiceProvider? services)
     {
         var compactionStrategy = new ContextWindowCompactionStrategy(
             maxContextWindowTokens: maxContextWindowTokens,
@@ -165,13 +174,13 @@ public sealed class HarnessAgent : DelegatingAIAgent
 
         ChatOptions chatOptions = BuildChatOptions(options, instructions, maxOutputTokens);
 
-        var compactionProvider = new CompactionProvider(compactionStrategy);
+        var compactionProvider = new CompactionProvider(compactionStrategy, loggerFactory: loggerFactory);
 
-        IEnumerable<AIContextProvider> contextProviders = BuildContextProviders(options);
+        IEnumerable<AIContextProvider> contextProviders = BuildContextProviders(options, loggerFactory);
 
         return chatClient
             .AsBuilder()
-            .UseFunctionInvocation(configure: options?.MaximumIterationsPerRequest is int maxIterations
+            .UseFunctionInvocation(loggerFactory, configure: options?.MaximumIterationsPerRequest is int maxIterations
                 ? ficc => ficc.MaximumIterationsPerRequest = maxIterations
                 : null)
             .UseMessageInjection()
@@ -189,7 +198,9 @@ public sealed class HarnessAgent : DelegatingAIAgent
                 RequirePerServiceCallChatHistoryPersistence = true,
                 WarnOnChatHistoryProviderConflict = false,
                 ThrowOnChatHistoryProviderConflict = false,
-            });
+            },
+            loggerFactory,
+            services);
     }
 
     private static ChatOptions BuildChatOptions(HarnessAgentOptions? options, string instructions, int maxOutputTokens)
@@ -215,7 +226,7 @@ public sealed class HarnessAgent : DelegatingAIAgent
         return result;
     }
 
-    private static List<AIContextProvider> BuildContextProviders(HarnessAgentOptions? options)
+    private static List<AIContextProvider> BuildContextProviders(HarnessAgentOptions? options, ILoggerFactory? loggerFactory)
     {
         var providers = new List<AIContextProvider>();
 
@@ -255,8 +266,8 @@ public sealed class HarnessAgent : DelegatingAIAgent
         if (options?.DisableAgentSkillsProvider is not true)
         {
             AgentSkillsProvider skillsProvider = options?.AgentSkillsSource is AgentSkillsSource source
-                ? new AgentSkillsProvider(source)
-                : new AgentSkillsProvider(Directory.GetCurrentDirectory());
+                ? new AgentSkillsProvider(source, loggerFactory: loggerFactory)
+                : new AgentSkillsProvider(Directory.GetCurrentDirectory(), loggerFactory: loggerFactory);
 
             providers.Add(skillsProvider);
         }

dotnet/src/Microsoft.Agents.AI.Hosting.AGUI.AspNetCore/AGUIEndpointRouteBuilderExtensions.cs

@@ -103,7 +103,16 @@ public static class AGUIEndpointRouteBuilderExtensions
         ArgumentNullException.ThrowIfNull(aiAgent);
 
         var agentSessionStore = endpoints.ServiceProvider.GetKeyedService<AgentSessionStore>(aiAgent.Name);
-        var hostAgent = new AIHostAgent(aiAgent, agentSessionStore ?? new NoopAgentSessionStore());
+
+        // Ensure that we have an IsolationKeyScopedAgentSessionStore registered.
+        var isolationKeyProvider = endpoints.ServiceProvider.GetService<SessionIsolationKeyProvider>();
+        if (agentSessionStore?.GetService<IsolationKeyScopedAgentSessionStore>() is null)
+        {
+            agentSessionStore ??= new NoopAgentSessionStore();
+            agentSessionStore = new IsolationKeyScopedAgentSessionStore(agentSessionStore, isolationKeyProvider, new() { Strict = isolationKeyProvider != null });
+        }
+
+        var hostAgent = new AIHostAgent(aiAgent, agentSessionStore);
 
         return endpoints.MapPost(pattern, async ([FromBody] RunAgentInput? input, HttpContext context, CancellationToken cancellationToken) =>
         {

dotnet/src/Microsoft.Agents.AI.Workflows.Declarative.Foundry/Microsoft.Agents.AI.Workflows.Declarative.Foundry.csproj

@@ -1,7 +1,7 @@
 <Project Sdk="Microsoft.NET.Sdk">
 
   <PropertyGroup>
-    <IsReleaseCandidate>true</IsReleaseCandidate>
+    <!-- Preview while Microsoft.Agents.AI.Foundry is preview (blocked by Azure.AI.Projects 2.1.0-beta). Flip to IsReleased=true once that ships stable. -->
     <NoWarn>$(NoWarn);MEAI001;OPENAI001</NoWarn>
   </PropertyGroup>
 

dotnet/src/Microsoft.Agents.AI.Workflows.Declarative.Mcp/Microsoft.Agents.AI.Workflows.Declarative.Mcp.csproj

@@ -1,7 +1,7 @@
 <Project Sdk="Microsoft.NET.Sdk">
 
   <PropertyGroup>
-    <IsReleaseCandidate>true</IsReleaseCandidate>
+    <IsReleased>true</IsReleased>
     <NoWarn>$(NoWarn);MEAI001;OPENAI001</NoWarn>
   </PropertyGroup>
 
@@ -13,9 +13,11 @@
 
   <Import Project="$(RepoRoot)/dotnet/nuget/nuget-package.props" />
 
-  <!-- Package not yet published to NuGet — disable baseline validation until first release -->
+  <!-- First Stable release after the RC milestone. Baseline against the latest 
+       published RC so package validation catches accidental breaking changes. 
+       Future releases should bump this to the previous stable version. -->
   <PropertyGroup>
-    <EnablePackageValidation>false</EnablePackageValidation>
+    <PackageValidationBaselineVersion>1.8.0-rc1</PackageValidationBaselineVersion>
   </PropertyGroup>
 
   <PropertyGroup>

dotnet/src/Microsoft.Agents.AI.Workflows.Declarative/Microsoft.Agents.AI.Workflows.Declarative.csproj

@@ -1,7 +1,7 @@
 <Project Sdk="Microsoft.NET.Sdk">
 
   <PropertyGroup>
-    <IsReleaseCandidate>true</IsReleaseCandidate>
+    <IsReleased>true</IsReleased>
     <NoWarn>$(NoWarn);MEAI001;OPENAI001</NoWarn>
   </PropertyGroup>
 
@@ -13,6 +13,13 @@
 
   <Import Project="$(RepoRoot)/dotnet/nuget/nuget-package.props" />
 
+  <!-- First Stable release after the RC milestone. Baseline against the latest 
+       published RC so package validation catches accidental breaking changes. 
+       Future releases should bump this to the previous stable version. -->
+  <PropertyGroup>
+    <PackageValidationBaselineVersion>1.8.0-rc1</PackageValidationBaselineVersion>
+  </PropertyGroup>
+
   <PropertyGroup>
     <!-- NuGet Package Settings -->
     <Title>Microsoft Agent Framework Declarative Workflows</Title>

dotnet/src/Microsoft.Agents.AI.Workflows.Declarative/ObjectModel/InvokeMcpToolExecutor.cs

@@ -27,6 +27,14 @@ internal sealed class InvokeMcpToolExecutor(
     WorkflowFormulaState state) :
     DeclarativeActionExecutor<InvokeMcpTool>(model, state)
 {
+    private const string ApprovalSnapshotStateKey = nameof(_approvalSnapshot);
+
+    /// <summary>
+    /// Snapshot of evaluated parameters at approval-request time.
+    /// Used to prevent TOCTOU attacks where state mutates during the approval window.
+    /// </summary>
+    private ApprovalSnapshot? _approvalSnapshot;
+
     /// <summary>
     /// Step identifiers for the MCP tool invocation workflow.
     /// </summary>
@@ -75,6 +83,10 @@ internal sealed class InvokeMcpToolExecutor(
 
         if (requireApproval)
         {
+            // Snapshot the evaluated parameters to prevent TOCTOU attacks.
+            // If state mutates during the approval window, the approved values are used on resume.
+            this._approvalSnapshot = new ApprovalSnapshot(serverUrl, serverLabel, toolName, arguments, connectionName);
+
             // Create tool call content for approval request.
             // Transport headers (e.g. Authorization) are intentionally excluded from the
             // approval event: they must not cross into the externally-surfaced approval request.
@@ -137,13 +149,14 @@ internal sealed class InvokeMcpToolExecutor(
             return;
         }
 
-        // Approved - now invoke the tool
-        string serverUrl = this.GetServerUrl();
-        string? serverLabel = this.GetServerLabel();
-        string toolName = this.GetToolName();
-        Dictionary<string, object?>? arguments = this.GetArguments();
+        // Approved - use the snapshot from approval-request time to prevent TOCTOU attacks.
+        // Headers are re-evaluated (they may contain auth secrets that should not be persisted).
+        string serverUrl = this._approvalSnapshot?.ServerUrl ?? this.GetServerUrl();
+        string? serverLabel = this._approvalSnapshot?.ServerLabel ?? this.GetServerLabel();
+        string toolName = this._approvalSnapshot?.ToolName ?? this.GetToolName();
+        Dictionary<string, object?>? arguments = this._approvalSnapshot?.Arguments ?? this.GetArguments();
         Dictionary<string, string>? headers = this.GetHeaders();
-        string? connectionName = this.GetConnectionName();
+        string? connectionName = this._approvalSnapshot?.ConnectionName ?? this.GetConnectionName();
 
         McpServerToolResultContent resultContent = await mcpToolHandler.InvokeToolAsync(
             serverUrl,
@@ -162,9 +175,33 @@ internal sealed class InvokeMcpToolExecutor(
     /// </summary>
     public async ValueTask CompleteAsync(IWorkflowContext context, ActionExecutorResult message, CancellationToken cancellationToken)
     {
+        // Clear the approval snapshot after successful completion.
+        this._approvalSnapshot = null;
+        await ClearSnapshotStateAsync(context, cancellationToken).ConfigureAwait(false);
+
         await context.RaiseCompletionEventAsync(this.Model, cancellationToken).ConfigureAwait(false);
     }
 
+    /// <inheritdoc/>
+    /// <remarks>
+    /// Persists the approval snapshot to workflow state so it survives checkpoint/restore cycles.
+    /// </remarks>
+    protected override async ValueTask OnCheckpointingAsync(IWorkflowContext context, CancellationToken cancellationToken = default)
+    {
+        await context.QueueStateUpdateAsync(ApprovalSnapshotStateKey, this._approvalSnapshot, null, cancellationToken).ConfigureAwait(false);
+        await base.OnCheckpointingAsync(context, cancellationToken).ConfigureAwait(false);
+    }
+
+    /// <inheritdoc/>
+    /// <remarks>
+    /// Restores the approval snapshot from workflow state after a checkpoint restore.
+    /// </remarks>
+    protected override async ValueTask OnCheckpointRestoredAsync(IWorkflowContext context, CancellationToken cancellationToken = default)
+    {
+        await base.OnCheckpointRestoredAsync(context, cancellationToken).ConfigureAwait(false);
+        this._approvalSnapshot = await context.ReadStateAsync<ApprovalSnapshot>(ApprovalSnapshotStateKey, null, cancellationToken).ConfigureAwait(false);
+    }
+
     private async ValueTask ProcessResultAsync(IWorkflowContext context, McpServerToolResultContent resultContent, CancellationToken cancellationToken)
     {
         bool autoSend = this.GetAutoSendValue();
@@ -365,4 +402,24 @@ internal sealed class InvokeMcpToolExecutor(
 
         return result;
     }
+
+    /// <summary>
+    /// Clears the persisted approval snapshot state after a successful tool invocation.
+    /// </summary>
+    private static async ValueTask ClearSnapshotStateAsync(IWorkflowContext context, CancellationToken cancellationToken)
+    {
+        await context.QueueStateUpdateAsync<ApprovalSnapshot?>(ApprovalSnapshotStateKey, null, null, cancellationToken).ConfigureAwait(false);
+    }
+
+    /// <summary>
+    /// Stores the evaluated parameters at approval-request time so that
+    /// <see cref="CaptureResponseAsync"/> uses the values the user reviewed,
+    /// even if <see cref="WorkflowFormulaState"/> mutates during the approval window.
+    /// </summary>
+    internal sealed record ApprovalSnapshot(
+        string ServerUrl,
+        string? ServerLabel,
+        string ToolName,
+        Dictionary<string, object?>? Arguments,
+        string? ConnectionName);
 }

dotnet/tests/Foundry.Hosting.IntegrationTests/scripts/it-bootstrap-agents.ps1

@@ -6,11 +6,11 @@
 .DESCRIPTION
   The IT fixture targets stable, scenario-keyed agent names (e.g. it-happy-path) and only
   manages versions on each test run. The agent itself must already exist AND its managed
-  identity must hold the Azure AI User role on the project scope, otherwise inbound
+  identity must hold the Foundry User role on the project scope, otherwise inbound
   inference calls fail with HTTP 500 PermissionDenied.
 
   This script idempotently creates each scenario agent (with a placeholder version) and
-  grants Azure AI User on the project to its managed identity. Re-run it safely; existing
+  grants Foundry User on the project to its managed identity. Re-run it safely; existing
   agents and role assignments are left in place.
 
 .PARAMETER ProjectEndpoint
@@ -135,20 +135,20 @@ foreach ($scenario in $Scenarios) {
             -Body $patchBody | Out-Null
     }
 
-    # 3. Grant Azure AI User on the project scope to the agent MI (idempotent).
+    # 3. Grant Foundry User on the project scope to the agent MI (idempotent).
     $existing = az role assignment list --assignee $principalId --scope $projectScope `
-        --query "[?roleDefinitionName=='Azure AI User']" 2>$null | ConvertFrom-Json
+        --query "[?roleDefinitionName=='Foundry User']" 2>$null | ConvertFrom-Json
     if ($existing) {
         Write-Host "  role already assigned"
     } else {
-        Write-Host "  granting Azure AI User..."
+        Write-Host "  granting Foundry User..."
         $maxAttempts = 12
         $granted = $false
         for ($i = 1; $i -le $maxAttempts; $i++) {
             $output = az role assignment create `
                 --assignee-object-id $principalId `
                 --assignee-principal-type ServicePrincipal `
-                --role 'Azure AI User' `
+                --role 'Foundry User' `
                 --scope $projectScope 2>&1
             if ($LASTEXITCODE -eq 0) {
                 $granted = $true

dotnet/tests/Microsoft.Agents.AI.DevUI.UnitTests/Properties/launchSettings.json

@@ -1,12 +0,0 @@
-{
-  "profiles": {
-    "Microsoft.Agents.AI.DevUI.UnitTests": {
-      "commandName": "Project",
-      "launchBrowser": true,
-      "environmentVariables": {
-        "ASPNETCORE_ENVIRONMENT": "Development"
-      },
-      "applicationUrl": "https://localhost:63009;http://localhost:63010"
-    }
-  }
-}

dotnet/tests/Microsoft.Agents.AI.Foundry.Hosting.UnitTests/OutputConverterTests.cs

@@ -704,6 +704,35 @@ public class OutputConverterTests
         Assert.Equal("[{\"id\":1}]", inner);
     }
 
+    // K-06e: Regression — the OutputItemFunctionToolCallOutput must have a populated Id
+    // and a matching wire id on the added/done events. The Foundry storage layer extracts
+    // a partition id from this field and throws "ID cannot be null or empty (Parameter 'id')"
+    // when it is missing.
+    [Fact]
+    public async Task ConvertUpdatesToEventsAsync_FunctionResult_OutputItemHasIdAsync()
+    {
+        var (stream, _) = CreateTestStream();
+        var update = new AgentResponseUpdate { Contents = [new FunctionResultContent("call_1", "sunny")] };
+
+        var events = new List<ResponseStreamEvent>();
+        await foreach (var evt in OutputConverter.ConvertUpdatesToEventsAsync(ToAsync(new[] { update }), stream))
+        {
+            events.Add(evt);
+        }
+
+        var added = Assert.Single(events.OfType<ResponseOutputItemAddedEvent>());
+        var done = Assert.Single(events.OfType<ResponseOutputItemDoneEvent>());
+
+        var addedOutput = Assert.IsType<OutputItemFunctionToolCallOutput>(added.Item);
+        var doneOutput = Assert.IsType<OutputItemFunctionToolCallOutput>(done.Item);
+
+        Assert.False(string.IsNullOrEmpty(addedOutput.Id));
+        Assert.False(string.IsNullOrEmpty(doneOutput.Id));
+        Assert.Equal(addedOutput.Id, doneOutput.Id);
+        Assert.Equal("call_1", addedOutput.CallId);
+        Assert.Equal("call_1", doneOutput.CallId);
+    }
+
     // L-01
     [Fact]
     public async Task ConvertUpdatesToEventsAsync_ExecutorInvokedEvent_EmitsWorkflowActionItemAsync()

dotnet/tests/Microsoft.Agents.AI.Harness.UnitTests/HarnessAgentTests.cs

@@ -9,6 +9,7 @@ using System.Threading.Tasks;
 using Microsoft.Agents.AI.Tools.Shell;
 #endif
 using Microsoft.Extensions.AI;
+using Microsoft.Extensions.Logging;
 using Moq;
 
 namespace Microsoft.Agents.AI.UnitTests;
@@ -1460,4 +1461,131 @@ public class HarnessAgentTests
 
     #endregion
 #endif
+
+    #region LoggerFactory and ServiceProvider
+
+    /// <summary>
+    /// Verify that the constructor succeeds when loggerFactory is provided.
+    /// </summary>
+    [Fact]
+    public void Constructor_SucceedsWithLoggerFactory()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var loggerFactory = new Mock<ILoggerFactory>().Object;
+
+        // Act
+        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), loggerFactory);
+
+        // Assert
+        Assert.NotNull(agent);
+    }
+
+    /// <summary>
+    /// Verify that the constructor succeeds when serviceProvider is provided.
+    /// </summary>
+    [Fact]
+    public void Constructor_SucceedsWithServiceProvider()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var services = new Mock<IServiceProvider>().Object;
+
+        // Act
+        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), services: services);
+
+        // Assert
+        Assert.NotNull(agent);
+    }
+
+    /// <summary>
+    /// Verify that the constructor succeeds when both loggerFactory and serviceProvider are provided.
+    /// </summary>
+    [Fact]
+    public void Constructor_SucceedsWithLoggerFactoryAndServiceProvider()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var loggerFactory = new Mock<ILoggerFactory>().Object;
+        var services = new Mock<IServiceProvider>().Object;
+
+        // Act
+        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), loggerFactory, services);
+
+        // Assert
+        Assert.NotNull(agent);
+    }
+
+    /// <summary>
+    /// Verify that AsHarnessAgent extension method accepts loggerFactory and serviceProvider.
+    /// </summary>
+    [Fact]
+    public void AsHarnessAgent_SucceedsWithLoggerFactoryAndServiceProvider()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var loggerFactory = new Mock<ILoggerFactory>().Object;
+        var services = new Mock<IServiceProvider>().Object;
+
+        // Act
+        var agent = chatClient.AsHarnessAgent(TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), loggerFactory, services);
+
+        // Assert
+        Assert.NotNull(agent);
+    }
+
+    /// <summary>
+    /// Verify that ILoggerFactory is threaded to downstream components by confirming CreateLogger is called.
+    /// </summary>
+    [Fact]
+    public void Constructor_LoggerFactoryIsUsedByDownstreamComponents()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var mockLoggerFactory = new Mock<ILoggerFactory>();
+        mockLoggerFactory
+            .Setup(lf => lf.CreateLogger(It.IsAny<string>()))
+            .Returns(new Mock<ILogger>().Object);
+
+        // Act — use options that leave CompactionProvider and AgentSkillsProvider enabled
+        var options = new HarnessAgentOptions
+        {
+            DisableToolApproval = true,
+            DisableOpenTelemetry = true,
+            DisableFileMemory = true,
+            DisableFileAccess = true,
+            DisableWebSearch = true,
+            DisableTodoProvider = true,
+            DisableAgentModeProvider = true,
+        };
+        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, options, mockLoggerFactory.Object);
+
+        // Assert — CreateLogger should have been called by one or more downstream components
+        Assert.NotNull(agent);
+        mockLoggerFactory.Verify(lf => lf.CreateLogger(It.IsAny<string>()), Times.AtLeastOnce());
+    }
+
+    /// <summary>
+    /// Verify that IServiceProvider is propagated through the agent pipeline by confirming
+    /// it is queried during agent construction.
+    /// </summary>
+    [Fact]
+    public void Constructor_ServiceProviderIsQueriedDuringBuild()
+    {
+        // Arrange
+        var chatClient = new Mock<IChatClient>().Object;
+        var mockServices = new Mock<IServiceProvider>();
+        mockServices
+            .Setup(sp => sp.GetService(It.IsAny<Type>()))
+            .Returns(null!);
+
+        // Act
+        var agent = new HarnessAgent(chatClient, TestMaxContextWindowTokens, TestMaxOutputTokens, CreateAllDisabledOptions(), services: mockServices.Object);
+
+        // Assert — the service provider should have been queried during pipeline construction
+        Assert.NotNull(agent);
+        mockServices.Verify(sp => sp.GetService(It.IsAny<Type>()), Times.AtLeastOnce());
+    }
+
+    #endregion
 }

dotnet/tests/Microsoft.Agents.AI.Hosting.A2A.UnitTests/Properties/launchSettings.json

@@ -1,12 +0,0 @@
-{
-  "profiles": {
-    "Microsoft.Agents.AI.Hosting.A2A.UnitTests": {
-      "commandName": "Project",
-      "launchBrowser": true,
-      "environmentVariables": {
-        "ASPNETCORE_ENVIRONMENT": "Development"
-      },
-      "applicationUrl": "https://localhost:52186;http://localhost:52187"
-    }
-  }
-}

dotnet/tests/Microsoft.Agents.AI.Hosting.OpenAI.UnitTests/Properties/launchSettings.json

@@ -1,12 +0,0 @@
-{
-  "profiles": {
-    "Microsoft.Agents.AI.Hosting.OpenAI.UnitTests": {
-      "commandName": "Project",
-      "launchBrowser": true,
-      "environmentVariables": {
-        "ASPNETCORE_ENVIRONMENT": "Development"
-      },
-      "applicationUrl": "https://localhost:60491;http://localhost:60492"
-    }
-  }
-}

dotnet/tests/Microsoft.Agents.AI.Workflows.Declarative.UnitTests/ObjectModel/InvokeMcpToolExecutorTest.cs

@@ -2,6 +2,7 @@
 
 using System.Collections.Generic;
 using System.Linq;
+using System.Reflection;
 using System.Threading;
 using System.Threading.Tasks;
 using Microsoft.Agents.AI.Workflows.Declarative.Events;
@@ -11,7 +12,9 @@ using Microsoft.Agents.AI.Workflows.Declarative.ObjectModel;
 using Microsoft.Agents.AI.Workflows.Declarative.PowerFx;
 using Microsoft.Agents.ObjectModel;
 using Microsoft.Extensions.AI;
+using Microsoft.PowerFx.Types;
 using Moq;
+using ApprovalSnapshot = Microsoft.Agents.AI.Workflows.Declarative.ObjectModel.InvokeMcpToolExecutor.ApprovalSnapshot;
 
 namespace Microsoft.Agents.AI.Workflows.Declarative.UnitTests.ObjectModel;
 
@@ -842,6 +845,313 @@ public sealed class InvokeMcpToolExecutorTest(ITestOutputHelper output) : Workfl
 
     #endregion
 
+    #region Approval Snapshot Security Tests
+
+    /// <summary>
+    /// Verifies that mutating the tool name variable after approval does not change
+    /// which tool is actually invoked. The originally-approved tool name must be used.
+    /// </summary>
+    [Fact]
+    public async Task InvokeMcpToolCaptureResponseUsesApprovedToolNameNotMutatedAsync()
+    {
+        // Arrange
+        const string ApprovedToolName = "safe_readonly_query";
+        const string MutatedToolName = "dangerous_admin_tool";
+
+        this.State.Set("TargetTool", FormulaValue.New(ApprovedToolName));
+        this.State.InitializeSystem();
+        this.State.Bind();
+
+        InvokeMcpTool model = this.CreateModelWithVariableToolName(
+            displayName: nameof(InvokeMcpToolCaptureResponseUsesApprovedToolNameNotMutatedAsync),
+            serverUrl: TestServerUrl,
+            variableName: "TargetTool");
+
+        string? capturedToolName = null;
+        Mock<IMcpToolHandler> mockProvider = new();
+        mockProvider.Setup(provider => provider.InvokeToolAsync(
+                It.IsAny<string>(),
+                It.IsAny<string?>(),
+                It.IsAny<string>(),
+                It.IsAny<IDictionary<string, object?>?>(),
+                It.IsAny<IDictionary<string, string>?>(),
+                It.IsAny<string?>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<string, string?, string, IDictionary<string, object?>?, IDictionary<string, string>?, string?, CancellationToken>(
+                (_, _, toolName, _, _, _, _) => capturedToolName = toolName)
+            .ReturnsAsync(new McpServerToolResultContent("capture-call-id")
+            {
+                Outputs = [new TextContent("result")]
+            });
+        MockAgentProvider mockAgentProvider = new();
+        InvokeMcpToolExecutor action = new(model, mockProvider.Object, mockAgentProvider.Object, this.State);
+
+        // Act - trigger ExecuteAsync to store the approval snapshot
+        Mock<IWorkflowContext> mockContext = CreateMockWorkflowContext();
+        await action.HandleAsync(new ActionExecutorResult(action.Id), mockContext.Object, CancellationToken.None);
+
+        // Simulate parallel branch mutating state during the approval window
+        this.State.Set("TargetTool", FormulaValue.New(MutatedToolName));
+        this.State.Bind();
+
+        // User clicks approve (they saw "safe_readonly_query" in the approval UI)
+        McpServerToolCallContent toolCall = new(action.Id, ApprovedToolName, TestServerUrl);
+        ToolApprovalRequestContent approvalRequest = new(action.Id, toolCall);
+        ToolApprovalResponseContent approvalResponse = approvalRequest.CreateResponse(approved: true);
+        ExternalInputResponse response = new(new ChatMessage(ChatRole.User, [approvalResponse]));
+
+        // Resume after approval
+        await action.CaptureResponseAsync(mockContext.Object, response, CancellationToken.None);
+
+        // Assert - the originally-approved tool name must be used, not the mutated one
+        Assert.NotNull(capturedToolName);
+        Assert.Equal(ApprovedToolName, capturedToolName);
+    }
+
+    /// <summary>
+    /// Verifies that mutating an argument variable after approval does not change
+    /// the arguments actually passed to the MCP tool. The originally-approved arguments must be used.
+    /// </summary>
+    [Fact]
+    public async Task InvokeMcpToolCaptureResponseUsesApprovedArgumentsNotMutatedAsync()
+    {
+        // Arrange
+        const string ApprovedQuery = "SELECT * FROM users LIMIT 10";
+        const string MutatedQuery = "DROP TABLE users CASCADE; --";
+
+        this.State.Set("SqlQuery", FormulaValue.New(ApprovedQuery));
+        this.State.InitializeSystem();
+        this.State.Bind();
+
+        InvokeMcpTool model = this.CreateModelWithVariableArgument(
+            displayName: nameof(InvokeMcpToolCaptureResponseUsesApprovedArgumentsNotMutatedAsync),
+            serverUrl: TestServerUrl,
+            toolName: TestToolName,
+            argumentKey: "query",
+            variableName: "SqlQuery");
+
+        IDictionary<string, object?>? capturedArguments = null;
+        Mock<IMcpToolHandler> mockProvider = new();
+        mockProvider.Setup(provider => provider.InvokeToolAsync(
+                It.IsAny<string>(),
+                It.IsAny<string?>(),
+                It.IsAny<string>(),
+                It.IsAny<IDictionary<string, object?>?>(),
+                It.IsAny<IDictionary<string, string>?>(),
+                It.IsAny<string?>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<string, string?, string, IDictionary<string, object?>?, IDictionary<string, string>?, string?, CancellationToken>(
+                (_, _, _, arguments, _, _, _) => capturedArguments = arguments)
+            .ReturnsAsync(new McpServerToolResultContent("capture-call-id")
+            {
+                Outputs = [new TextContent("result")]
+            });
+        MockAgentProvider mockAgentProvider = new();
+        InvokeMcpToolExecutor action = new(model, mockProvider.Object, mockAgentProvider.Object, this.State);
+
+        // Act - trigger ExecuteAsync to store the approval snapshot
+        Mock<IWorkflowContext> mockContext = CreateMockWorkflowContext();
+        await action.HandleAsync(new ActionExecutorResult(action.Id), mockContext.Object, CancellationToken.None);
+
+        // Simulate parallel branch mutating state during the approval window
+        this.State.Set("SqlQuery", FormulaValue.New(MutatedQuery));
+        this.State.Bind();
+
+        // User clicks approve
+        McpServerToolCallContent toolCall = new(action.Id, TestToolName, TestServerUrl);
+        ToolApprovalRequestContent approvalRequest = new(action.Id, toolCall);
+        ToolApprovalResponseContent approvalResponse = approvalRequest.CreateResponse(approved: true);
+        ExternalInputResponse response = new(new ChatMessage(ChatRole.User, [approvalResponse]));
+
+        // Resume after approval
+        await action.CaptureResponseAsync(mockContext.Object, response, CancellationToken.None);
+
+        // Assert - the originally-approved argument must be used, not the mutated one
+        Assert.NotNull(capturedArguments);
+        Assert.Equal(ApprovedQuery, capturedArguments["query"]?.ToString());
+    }
+
+    /// <summary>
+    /// Verifies that mutating the server URL variable after approval does not redirect
+    /// the MCP tool call to a different server. The originally-approved server URL must be used.
+    /// </summary>
+    [Fact]
+    public async Task InvokeMcpToolCaptureResponseUsesApprovedServerUrlNotMutatedAsync()
+    {
+        // Arrange
+        const string ApprovedServerUrl = "https://internal-mcp.corp";
+        const string MutatedServerUrl = "https://attacker.evil/steal";
+
+        this.State.Set("McpEndpoint", FormulaValue.New(ApprovedServerUrl));
+        this.State.InitializeSystem();
+        this.State.Bind();
+
+        InvokeMcpTool model = this.CreateModelWithVariableServerUrl(
+            displayName: nameof(InvokeMcpToolCaptureResponseUsesApprovedServerUrlNotMutatedAsync),
+            variableName: "McpEndpoint",
+            toolName: TestToolName);
+
+        string? capturedServerUrl = null;
+        Mock<IMcpToolHandler> mockProvider = new();
+        mockProvider.Setup(provider => provider.InvokeToolAsync(
+                It.IsAny<string>(),
+                It.IsAny<string?>(),
+                It.IsAny<string>(),
+                It.IsAny<IDictionary<string, object?>?>(),
+                It.IsAny<IDictionary<string, string>?>(),
+                It.IsAny<string?>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<string, string?, string, IDictionary<string, object?>?, IDictionary<string, string>?, string?, CancellationToken>(
+                (serverUrl, _, _, _, _, _, _) => capturedServerUrl = serverUrl)
+            .ReturnsAsync(new McpServerToolResultContent("capture-call-id")
+            {
+                Outputs = [new TextContent("result")]
+            });
+        MockAgentProvider mockAgentProvider = new();
+        InvokeMcpToolExecutor action = new(model, mockProvider.Object, mockAgentProvider.Object, this.State);
+
+        // Act - trigger ExecuteAsync to store the approval snapshot
+        Mock<IWorkflowContext> mockContext = CreateMockWorkflowContext();
+        await action.HandleAsync(new ActionExecutorResult(action.Id), mockContext.Object, CancellationToken.None);
+
+        // Simulate parallel branch mutating state during the approval window
+        this.State.Set("McpEndpoint", FormulaValue.New(MutatedServerUrl));
+        this.State.Bind();
+
+        // User clicks approve
+        McpServerToolCallContent toolCall = new(action.Id, TestToolName, ApprovedServerUrl);
+        ToolApprovalRequestContent approvalRequest = new(action.Id, toolCall);
+        ToolApprovalResponseContent approvalResponse = approvalRequest.CreateResponse(approved: true);
+        ExternalInputResponse response = new(new ChatMessage(ChatRole.User, [approvalResponse]));
+
+        // Resume after approval
+        await action.CaptureResponseAsync(mockContext.Object, response, CancellationToken.None);
+
+        // Assert - the originally-approved server URL must be used, not the mutated one
+        Assert.NotNull(capturedServerUrl);
+        Assert.Equal(ApprovedServerUrl, capturedServerUrl);
+    }
+
+    /// <summary>
+    /// Verifies that the approval snapshot survives a checkpoint/restore cycle.
+    /// After restore, the originally-approved tool name must still be used even if state was mutated.
+    /// </summary>
+    [Fact]
+    public async Task InvokeMcpToolCaptureResponseUsesSnapshotAfterCheckpointRestoreAsync()
+    {
+        // Arrange
+        const string ApprovedToolName = "safe_readonly_query";
+        const string MutatedToolName = "dangerous_admin_tool";
+
+        this.State.Set("TargetTool", FormulaValue.New(ApprovedToolName));
+        this.State.InitializeSystem();
+        this.State.Bind();
+
+        InvokeMcpTool model = this.CreateModelWithVariableToolName(
+            displayName: nameof(InvokeMcpToolCaptureResponseUsesSnapshotAfterCheckpointRestoreAsync),
+            serverUrl: TestServerUrl,
+            variableName: "TargetTool");
+
+        string? capturedToolName = null;
+        Mock<IMcpToolHandler> mockProvider = new();
+        mockProvider.Setup(provider => provider.InvokeToolAsync(
+                It.IsAny<string>(),
+                It.IsAny<string?>(),
+                It.IsAny<string>(),
+                It.IsAny<IDictionary<string, object?>?>(),
+                It.IsAny<IDictionary<string, string>?>(),
+                It.IsAny<string?>(),
+                It.IsAny<CancellationToken>()))
+            .Callback<string, string?, string, IDictionary<string, object?>?, IDictionary<string, string>?, string?, CancellationToken>(
+                (_, _, toolName, _, _, _, _) => capturedToolName = toolName)
+            .ReturnsAsync(new McpServerToolResultContent("capture-call-id")
+            {
+                Outputs = [new TextContent("result")]
+            });
+        MockAgentProvider mockAgentProvider = new();
+        InvokeMcpToolExecutor action = new(model, mockProvider.Object, mockAgentProvider.Object, this.State);
+
+        // Act - trigger ExecuteAsync to store the approval snapshot
+        Mock<IWorkflowContext> mockContext = CreateMockWorkflowContextWithStateStore();
+        await action.HandleAsync(new ActionExecutorResult(action.Id), mockContext.Object, CancellationToken.None);
+
+        // Simulate checkpoint: persist to state store
+        await InvokeProtectedMethodAsync(action, "OnCheckpointingAsync", mockContext.Object, CancellationToken.None);
+
+        // Simulate restore on a "new" executor instance by clearing the in-memory field via reflection
+        // (In production, a new executor instance would be created with _approvalSnapshot == null)
+        typeof(InvokeMcpToolExecutor)
+            .GetField("_approvalSnapshot", BindingFlags.NonPublic | BindingFlags.Instance)!
+            .SetValue(action, null);
+
+        // Restore from state store
+        await InvokeProtectedMethodAsync(action, "OnCheckpointRestoredAsync", mockContext.Object, CancellationToken.None);
+
+        // Mutate state after restore (simulating parallel branch)
+        this.State.Set("TargetTool", FormulaValue.New(MutatedToolName));
+        this.State.Bind();
+
+        // User clicks approve
+        McpServerToolCallContent toolCall = new(action.Id, ApprovedToolName, TestServerUrl);
+        ToolApprovalRequestContent approvalRequest = new(action.Id, toolCall);
+        ToolApprovalResponseContent approvalResponse = approvalRequest.CreateResponse(approved: true);
+        ExternalInputResponse response = new(new ChatMessage(ChatRole.User, [approvalResponse]));
+
+        // Resume after approval
+        await action.CaptureResponseAsync(mockContext.Object, response, CancellationToken.None);
+
+        // Assert - the originally-approved tool name must be used, not the mutated one
+        Assert.NotNull(capturedToolName);
+        Assert.Equal(ApprovedToolName, capturedToolName);
+    }
+
+    private static Mock<IWorkflowContext> CreateMockWorkflowContext()
+    {
+        Mock<IWorkflowContext> mockContext = new();
+        mockContext.Setup(c => c.AddEventAsync(It.IsAny<WorkflowEvent>(), It.IsAny<CancellationToken>()))
+            .Returns(default(ValueTask));
+        mockContext.Setup(c => c.QueueStateUpdateAsync(It.IsAny<string>(), It.IsAny<object?>(), It.IsAny<string?>(), It.IsAny<CancellationToken>()))
+            .Returns(default(ValueTask));
+        mockContext.Setup(c => c.SendMessageAsync(It.IsAny<object>(), It.IsAny<string?>(), It.IsAny<CancellationToken>()))
+            .Returns(default(ValueTask));
+        return mockContext;
+    }
+
+    /// <summary>
+    /// Creates a mock workflow context that actually stores state values (for checkpoint/restore tests).
+    /// </summary>
+    private static Mock<IWorkflowContext> CreateMockWorkflowContextWithStateStore()
+    {
+        Dictionary<string, object?> stateStore = new();
+        Mock<IWorkflowContext> mockContext = new();
+        mockContext.Setup(c => c.AddEventAsync(It.IsAny<WorkflowEvent>(), It.IsAny<CancellationToken>()))
+            .Returns(default(ValueTask));
+        mockContext.Setup(c => c.QueueStateUpdateAsync(It.IsAny<string>(), It.IsAny<ApprovalSnapshot?>(), It.IsAny<string?>(), It.IsAny<CancellationToken>()))
+            .Callback<string, ApprovalSnapshot?, string?, CancellationToken>((key, value, _, _) => stateStore[key] = value)
+            .Returns(default(ValueTask));
+        mockContext.Setup(c => c.SendMessageAsync(It.IsAny<object>(), It.IsAny<string?>(), It.IsAny<CancellationToken>()))
+            .Returns(default(ValueTask));
+        mockContext.Setup(c => c.ReadStateAsync<ApprovalSnapshot>(It.IsAny<string>(), It.IsAny<string?>(), It.IsAny<CancellationToken>()))
+            .Returns<string, string?, CancellationToken>((key, _, _) =>
+                new ValueTask<ApprovalSnapshot?>(stateStore.TryGetValue(key, out object? val) ? val as ApprovalSnapshot : null));
+        mockContext.Setup(c => c.ReadStateKeysAsync(It.IsAny<string?>(), It.IsAny<CancellationToken>()))
+            .ReturnsAsync(new HashSet<string>());
+        return mockContext;
+    }
+
+    /// <summary>
+    /// Invokes a protected method on an executor via reflection (for testing checkpoint hooks).
+    /// </summary>
+    private static async ValueTask InvokeProtectedMethodAsync(InvokeMcpToolExecutor action, string methodName, IWorkflowContext context, CancellationToken cancellationToken)
+    {
+        MethodInfo method = typeof(InvokeMcpToolExecutor)
+            .GetMethod(methodName, BindingFlags.NonPublic | BindingFlags.Instance)!;
+        ValueTask result = (ValueTask)method.Invoke(action, [context, cancellationToken])!;
+        await result.ConfigureAwait(false);
+    }
+
+    #endregion
+
     #region CompleteAsync Tests
 
     [Fact]
@@ -951,6 +1261,50 @@ public sealed class InvokeMcpToolExecutorTest(ITestOutputHelper output) : Workfl
         return AssignParent<InvokeMcpTool>(builder);
     }
 
+    private InvokeMcpTool CreateModelWithVariableToolName(string displayName, string serverUrl, string variableName)
+    {
+        InvokeMcpTool.Builder builder = new()
+        {
+            Id = this.CreateActionId(),
+            DisplayName = this.FormatDisplayName(displayName),
+            ServerUrl = new StringExpression.Builder(StringExpression.Literal(serverUrl)),
+            ToolName = new StringExpression.Builder(
+                StringExpression.Variable(PropertyPath.TopicVariable(variableName))),
+            RequireApproval = new BoolExpression.Builder(BoolExpression.Literal(true)),
+        };
+        return AssignParent<InvokeMcpTool>(builder);
+    }
+
+    private InvokeMcpTool CreateModelWithVariableArgument(
+        string displayName, string serverUrl, string toolName, string argumentKey, string variableName)
+    {
+        InvokeMcpTool.Builder builder = new()
+        {
+            Id = this.CreateActionId(),
+            DisplayName = this.FormatDisplayName(displayName),
+            ServerUrl = new StringExpression.Builder(StringExpression.Literal(serverUrl)),
+            ToolName = new StringExpression.Builder(StringExpression.Literal(toolName)),
+            RequireApproval = new BoolExpression.Builder(BoolExpression.Literal(true)),
+        };
+        builder.Arguments.Add(argumentKey,
+            ValueExpression.Variable(PropertyPath.TopicVariable(variableName)));
+        return AssignParent<InvokeMcpTool>(builder);
+    }
+
+    private InvokeMcpTool CreateModelWithVariableServerUrl(string displayName, string variableName, string toolName)
+    {
+        InvokeMcpTool.Builder builder = new()
+        {
+            Id = this.CreateActionId(),
+            DisplayName = this.FormatDisplayName(displayName),
+            ServerUrl = new StringExpression.Builder(
+                StringExpression.Variable(PropertyPath.TopicVariable(variableName))),
+            ToolName = new StringExpression.Builder(StringExpression.Literal(toolName)),
+            RequireApproval = new BoolExpression.Builder(BoolExpression.Literal(true)),
+        };
+        return AssignParent<InvokeMcpTool>(builder);
+    }
+
     #endregion
 
     #region Mock MCP Tool Provider

python/PACKAGE_STATUS.md

@@ -27,7 +27,7 @@ Status is grouped into these buckets:
 | `agent-framework-claude` | `python/packages/claude` | `beta` |
 | `agent-framework-copilotstudio` | `python/packages/copilotstudio` | `beta` |
 | `agent-framework-core` | `python/packages/core` | `released` |
-| `agent-framework-declarative` | `python/packages/declarative` | `beta` |
+| `agent-framework-declarative` | `python/packages/declarative` | `rc` |
 | `agent-framework-devui` | `python/packages/devui` | `beta` |
 | `agent-framework-durabletask` | `python/packages/durabletask` | `beta` |
 | `agent-framework-foundry` | `python/packages/foundry` | `released` |
@@ -58,6 +58,13 @@ listed below.
 
 ### Experimental features
 
+#### `DECLARATIVE_AGENTS`
+
+- `agent-framework-declarative`: declarative agent loading APIs from
+  `agent_framework_declarative`, including `AgentFactory`,
+  `DeclarativeLoaderError`, `ProviderLookupError`, and `ProviderTypeMapping`
+  from `agent_framework_declarative/_loader.py`
+
 #### `EVALS`
 
 - `agent-framework-core`: exported evaluation APIs from `agent_framework`, including

python/packages/a2a/agent_framework_a2a/_a2a_executor.py

@@ -287,9 +287,7 @@ class A2AExecutor(AgentExecutor):
                     artifact_id=artifact_id,
                     metadata=metadata,
                     append=(
-                        True
-                        if streamed_artifact_ids is not None and artifact_id in streamed_artifact_ids
-                        else None
+                        True if streamed_artifact_ids is not None and artifact_id in streamed_artifact_ids else None
                     ),
                 )
                 if artifact_id and streamed_artifact_ids is not None:

python/packages/anthropic/agent_framework_anthropic/_chat_client.py

@@ -803,6 +803,15 @@ class RawAnthropicClient(
                     }
                     a_content.append(mcp_result)
                 case "text_reasoning":
+                    if content.text is None:
+                        if (
+                            content.protected_data
+                            and a_content
+                            and a_content[-1].get("type") == "thinking"
+                            and "signature" not in a_content[-1]
+                        ):
+                            a_content[-1]["signature"] = content.protected_data
+                        continue
                     thinking_block: dict[str, Any] = {"type": "thinking", "thinking": content.text}
                     if content.protected_data:
                         thinking_block["signature"] = content.protected_data

python/packages/anthropic/tests/test_anthropic_client.py

@@ -485,6 +485,48 @@ def test_prepare_message_for_anthropic_text_reasoning_with_signature(
     assert result["content"][0]["signature"] == "sig_abc123"
 
 
+def test_prepare_message_for_anthropic_attaches_signature_only_reasoning(
+    mock_anthropic_client: MagicMock,
+) -> None:
+    client = create_test_anthropic_client(mock_anthropic_client)
+    message = Message(
+        role="assistant",
+        contents=[
+            Content.from_text_reasoning(text="Let me think about this..."),
+            Content.from_text_reasoning(text=None, protected_data="sig_abc123"),
+        ],
+    )
+
+    result = client._prepare_message_for_anthropic(message)
+
+    assert result["content"] == [
+        {"type": "thinking", "thinking": "Let me think about this...", "signature": "sig_abc123"}
+    ]
+
+
+def test_prepare_message_for_anthropic_skips_orphan_signature_only_reasoning(
+    mock_anthropic_client: MagicMock,
+) -> None:
+    client = create_test_anthropic_client(mock_anthropic_client)
+    message = Message(
+        role="assistant",
+        contents=[
+            Content.from_text_reasoning(text=None, protected_data="sig_abc123"),
+            Content.from_function_call(
+                call_id="call_123",
+                name="get_weather",
+                arguments={"location": "San Francisco"},
+            ),
+        ],
+    )
+
+    result = client._prepare_message_for_anthropic(message)
+
+    assert len(result["content"]) == 1
+    assert result["content"][0]["type"] == "tool_use"
+    assert result["content"][0]["id"] == "call_123"
+
+
 def test_prepare_message_for_anthropic_mcp_server_tool_call(
     mock_anthropic_client: MagicMock,
 ) -> None:

python/packages/bedrock/agent_framework_bedrock/_chat_client.py

@@ -4,6 +4,7 @@
 from __future__ import annotations
 
 import asyncio
+import copy
 import json
 import logging
 import sys
@@ -36,6 +37,7 @@ from agent_framework.observability import ChatTelemetryLayer
 from boto3.session import Session as Boto3Session
 from botocore.client import BaseClient
 from botocore.config import Config as BotoConfig
+from botocore.exceptions import ClientError
 from pydantic import BaseModel
 
 if sys.version_info >= (3, 13):
@@ -115,13 +117,20 @@ class BedrockChatOptions(ChatOptions[ResponseModelT], Generic[ResponseModelT], t
             translates to ``toolConfig.tools``.
         tool_choice: How the model should use tools,
             translates to ``toolConfig.toolChoice``.
+        response_format: Structured output format. Accepts a Pydantic BaseModel
+            subclass or an OpenAI-style dict schema
+            (``{"json_schema": {"name": ..., "schema": ...}}``).
+            When provided, the Converse API request includes
+            ``outputConfig.textFormat`` with the schema serialized as a JSON
+            string. ``ChatResponse.value`` will be populated with the parsed
+            model instance. Only supported on models that support
+            ``outputConfig.textFormat``. Unsupported models raise a ValueError.
 
         # Options not supported in Bedrock Converse API:
         seed: Not supported.
         frequency_penalty: Not supported.
         presence_penalty: Not supported.
         allow_multiple_tool_calls: Not supported (models handle parallel calls automatically).
-        response_format: Not directly supported (use model-specific prompting).
         user: Not supported.
         store: Not supported.
         logit_bias: Not supported.
@@ -161,9 +170,6 @@ class BedrockChatOptions(ChatOptions[ResponseModelT], Generic[ResponseModelT], t
     allow_multiple_tool_calls: None  # type: ignore[misc]
     """Not supported. Bedrock models handle parallel tool calls automatically."""
 
-    response_format: None  # type: ignore[misc]
-    """Not directly supported. Use model-specific prompting for JSON output."""
-
     user: None  # type: ignore[misc]
     """Not supported in Bedrock Converse API."""
 
@@ -324,10 +330,28 @@ class BedrockChatClient(
         return Boto3Session(**session_kwargs)
 
     def _invoke_converse(self, request: Mapping[str, Any]) -> dict[str, Any]:
-        response = self._bedrock_client.converse(**request)
-        if not isinstance(response, Mapping):
-            raise ChatClientInvalidResponseException("Bedrock converse response must be a mapping.")
-        return response
+        try:
+            response = self._bedrock_client.converse(**request)
+            if not isinstance(response, Mapping):
+                raise ChatClientInvalidResponseException("Bedrock converse response must be a mapping.")
+            return response
+        except ClientError as e:
+            error_details = e.response.get("Error", {})
+            error_code = error_details.get("Code", "")
+            error_message = error_details.get("Message", "")
+            # "outputConfig" in error_message catches cases where Bedrock explicitly
+            # rejects the outputConfig field (unsupported model). Other ValidationExceptions
+            # (e.g. malformed schema shape, invalid property values) will not mention
+            # "outputConfig" and will bubble up as raw ClientError without being misdiagnosed.
+            if error_code == "ValidationException" and (
+                "outputconfig" in error_message.lower() or "outputconfig" in str(e).lower()
+            ):
+                raise ValueError(
+                    f"Model '{self.model}' does not support structured output via outputConfig.textFormat. "
+                    "Check the model's Bedrock Converse outputConfig/textFormat support. "
+                    f"AWS error Code: {error_code}. AWS error Message: {error_message}"
+                ) from e
+            raise
 
     @override
     def _inner_get_response(
@@ -344,7 +368,7 @@ class BedrockChatClient(
             # Streaming mode - simulate streaming by yielding a single update
             async def _stream() -> AsyncIterable[ChatResponseUpdate]:
                 response = await asyncio.to_thread(self._invoke_converse, request)
-                parsed_response = self._process_converse_response(response)
+                parsed_response = self._process_converse_response(response, options)
                 contents = list(parsed_response.messages[0].contents if parsed_response.messages else [])
                 if parsed_response.usage_details:
                     contents.append(Content.from_usage(usage_details=parsed_response.usage_details))  # type: ignore[arg-type]
@@ -360,12 +384,12 @@ class BedrockChatClient(
                     raw_representation=parsed_response.raw_representation,
                 )
 
-            return self._build_response_stream(_stream())
+            return self._build_response_stream(_stream(), response_format=options.get("response_format"))
 
         # Non-streaming mode
         async def _get_response() -> ChatResponse:
             raw_response = await asyncio.to_thread(self._invoke_converse, request)
-            return self._process_converse_response(raw_response)
+            return self._process_converse_response(raw_response, options)
 
         return _get_response()
 
@@ -430,6 +454,9 @@ class BedrockChatClient(
         if tool_config:
             run_options["toolConfig"] = tool_config
 
+        if output_config := self._prepare_output_config(options.get("response_format")):
+            run_options["outputConfig"] = output_config
+
         return run_options
 
     def _prepare_bedrock_messages(
@@ -628,7 +655,9 @@ class BedrockChatClient(
     def _generate_tool_call_id() -> str:
         return f"tool-call-{uuid4().hex}"
 
-    def _process_converse_response(self, response: dict[str, Any]) -> ChatResponse:
+    def _process_converse_response(
+        self, response: dict[str, Any], options: Mapping[str, Any] | None = None
+    ) -> ChatResponse:
         """Convert Bedrock Converse API response to ChatResponse."""
         output = response.get("output") or {}
         message = output.get("message") or {}
@@ -646,6 +675,7 @@ class BedrockChatClient(
             usage_details=usage_details,
             model=model,
             finish_reason=finish_reason,
+            response_format=options.get("response_format") if options else None,
             raw_representation=response,
         )
 
@@ -728,6 +758,108 @@ class BedrockChatClient(
             return None
         return FINISH_REASON_MAP.get(reason.lower())
 
+    def _prepare_output_config(self, response_format: Any | None) -> dict[str, Any] | None:
+        """Convert response_format into the AWS Bedrock outputConfig wire format.
+
+        Args:
+            response_format: A Pydantic model class or a dict schema, or None.
+
+        Returns:
+            A dict for the Converse API ``outputConfig`` parameter, or None if
+            response_format is not set.
+        """
+        if response_format is None:
+            return None
+
+        if isinstance(response_format, Mapping):
+            if "json_schema" in response_format:
+                # Shape A — OpenAI-style wrapper
+                json_schema_config = response_format["json_schema"]
+                schema_src = json_schema_config.get("schema", {})
+                name = json_schema_config.get("name", "output_schema")
+            elif "schema" in response_format:
+                # Shape B — inner shape directly {"name": ..., "schema": ...}
+                schema_src = response_format["schema"]
+                name = response_format.get("name", "output_schema")
+            else:
+                # Shape C — assume entire dict is the raw schema
+                logger.warning(
+                    "response_format dict has no 'json_schema' or 'schema' key; "
+                    "treating entire dict as raw JSON schema."
+                )
+                schema_src = dict(response_format)
+                name = "output_schema"
+
+            if isinstance(schema_src, str):
+                schema_src = json.loads(schema_src)
+            schema = copy.deepcopy(schema_src)
+        else:
+            if not isinstance(response_format, type) or not issubclass(response_format, BaseModel):
+                raise TypeError(
+                    "response_format must be None, a dict JSON schema, "
+                    "or a Pydantic BaseModel subclass."
+                )
+            # response_format is a Pydantic model class
+            schema = response_format.model_json_schema()
+            name = response_format.__name__
+
+        self._set_additional_properties_false(schema)
+
+        json_schema: dict[str, Any] = {
+            "name": name,
+            "schema": json.dumps(schema),
+        }
+
+        description = getattr(response_format, "__doc__", None) if not isinstance(response_format, Mapping) else None
+        if description and isinstance(description, str) and description.strip():
+            json_schema["description"] = description.strip()
+
+        return {
+            "textFormat": {
+                "type": "json_schema",
+                "structure": {
+                    "jsonSchema": json_schema
+                },
+            }
+        }
+
+    def _set_additional_properties_false(self, schema: dict[str, Any]) -> None:
+        """Recursively set additionalProperties: false on all object types in a JSON schema.
+
+        AWS requires strict schema enforcement. This mirrors the approach used by
+        AnthropicChatClient._prepare_response_format().
+
+        Args:
+            schema: The JSON schema dict to modify in-place.
+        """
+        visited: set[int] = set()
+
+        def walk(node: Any) -> None:
+            if isinstance(node, dict):
+                node_id = id(node)
+                if node_id in visited:
+                    return
+                visited.add(node_id)
+                if node.get("type") == "object" or (
+                    "properties" in node and "type" not in node
+                ):
+                    existing = node.get("additionalProperties")
+                    if existing is None or existing is True:
+                        node["additionalProperties"] = False
+                for value in node.values():
+                    if isinstance(value, (dict, list)):
+                        walk(value)
+            elif isinstance(node, list):
+                node_id = id(node)
+                if node_id in visited:
+                    return
+                visited.add(node_id)
+                for item in node:
+                    if isinstance(item, (dict, list)):
+                        walk(item)
+
+        walk(schema)
+
     def service_url(self) -> str:
         """Returns the service URL for the Bedrock runtime in the configured AWS region.
 

python/packages/bedrock/tests/test_bedrock_structured_output.py

@@ -0,0 +1,382 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+from __future__ import annotations
+
+import copy
+import json
+from typing import Any
+from unittest.mock import patch
+
+import pytest
+from agent_framework import Content, Message
+from botocore.exceptions import ClientError
+from pydantic import BaseModel
+
+from agent_framework_bedrock import BedrockChatClient
+
+# region Test models
+
+
+class WeatherReport(BaseModel):
+    city: str
+    temperature: float
+    summary: str
+
+
+class NestedAddress(BaseModel):
+    street: str
+    city: str
+    zip_code: str
+
+
+class Person(BaseModel):
+    name: str
+    age: int
+    address: NestedAddress
+
+
+# endregion
+
+
+# region Helpers
+
+
+class _StubBedrockRuntime:
+    """Stub that records calls and returns a canned response."""
+
+    def __init__(self, response_text: str = "Bedrock says hi") -> None:
+        self.calls: list[dict[str, Any]] = []
+        self._response_text = response_text
+
+    def converse(self, **kwargs: Any) -> dict[str, Any]:
+        self.calls.append(kwargs)
+        return {
+            "modelId": kwargs["modelId"],
+            "responseId": "resp-structured",
+            "usage": {"inputTokens": 10, "outputTokens": 20, "totalTokens": 30},
+            "output": {
+                "completionReason": "end_turn",
+                "message": {
+                    "id": "msg-structured",
+                    "role": "assistant",
+                    "content": [{"text": self._response_text}],
+                },
+            },
+        }
+
+
+def _make_client(response_text: str = "Bedrock says hi") -> tuple[BedrockChatClient, _StubBedrockRuntime]:
+    stub = _StubBedrockRuntime(response_text)
+    client = BedrockChatClient(
+        model="us.anthropic.claude-haiku-4-5-v1:0",
+        region="us-east-1",
+        client=stub,
+    )
+    return client, stub
+
+
+def _user_messages() -> list[Message]:
+    return [Message(role="user", contents=[Content.from_text(text="Give me a weather report")])]
+
+
+# endregion
+
+
+# region Tests
+
+
+def test_prepare_output_config_correct_wire_shape() -> None:
+    """_prepare_output_config(WeatherReport) must produce the correct
+    textFormat → structure → jsonSchema shape with type: 'json_schema'."""
+    client, _ = _make_client()
+
+    output_config = client._prepare_output_config(WeatherReport)
+
+    assert output_config is not None
+    text_format = output_config["textFormat"]
+    assert text_format["type"] == "json_schema"
+    assert "structure" in text_format
+    json_schema = text_format["structure"]["jsonSchema"]
+    assert json_schema["name"] == "WeatherReport"
+    assert "schema" in json_schema
+
+
+def test_prepare_output_config_schema_is_json_string() -> None:
+    """The schema value inside jsonSchema must be a JSON string, not a dict."""
+    client, _ = _make_client()
+
+    output_config = client._prepare_output_config(WeatherReport)
+
+    assert output_config is not None
+    schema_value = output_config["textFormat"]["structure"]["jsonSchema"]["schema"]
+    assert isinstance(schema_value, str), f"Expected str, got {type(schema_value)}"
+    # Verify it's valid JSON
+    parsed = json.loads(schema_value)
+    assert isinstance(parsed, dict)
+    assert parsed["type"] == "object"
+
+
+def test_additional_properties_false_set_recursively() -> None:
+    """additionalProperties: false must be set on all nested object types."""
+    client, _ = _make_client()
+
+    output_config = client._prepare_output_config(Person)
+
+    assert output_config is not None
+    schema_str = output_config["textFormat"]["structure"]["jsonSchema"]["schema"]
+    schema = json.loads(schema_str)
+
+    # Top-level object
+    assert schema.get("additionalProperties") is False
+
+    # Check $defs for NestedAddress
+    defs = schema.get("$defs", {})
+    assert "NestedAddress" in defs, "Expected NestedAddress to be present in $defs"
+    assert defs["NestedAddress"].get("additionalProperties") is False, (
+        "Expected additionalProperties=False on nested NestedAddress schema"
+    )
+
+
+def test_no_output_config_when_response_format_none() -> None:
+    """When response_format is None, no outputConfig key should appear in the request."""
+    client, stub = _make_client()
+    messages = _user_messages()
+
+    request = client._prepare_options(messages, {"max_tokens": 100})
+
+    assert "outputConfig" not in request, (
+        f"outputConfig should not be present when response_format is None, got: {request.get('outputConfig')}"
+    )
+
+
+async def test_chat_response_value_populated() -> None:
+    """After a mocked response with response_format, .value should be a populated Pydantic model."""
+    json_response = json.dumps({"city": "Seattle", "temperature": 72.5, "summary": "Sunny and warm"})
+    client, stub = _make_client(response_text=json_response)
+    messages = _user_messages()
+
+    response = await client.get_response(
+        messages=messages,
+        options={"max_tokens": 100, "response_format": WeatherReport},
+    )
+
+    assert response.text == json_response
+    assert response.value is not None
+    assert isinstance(response.value, WeatherReport)
+    assert response.value.city == "Seattle"
+    assert response.value.temperature == 72.5
+    assert response.value.summary == "Sunny and warm"
+
+    # Verify outputConfig was sent to the API
+    assert len(stub.calls) == 1
+    api_request = stub.calls[0]
+    assert "outputConfig" in api_request
+    assert api_request["outputConfig"]["textFormat"]["type"] == "json_schema"
+
+
+def test_dict_schema_response_format() -> None:
+    """_prepare_output_config should work when response_format is a dict, not just a Pydantic class."""
+    client, _ = _make_client()
+
+    dict_schema = {
+        "json_schema": {
+            "name": "weather_output",
+            "schema": {
+                "type": "object",
+                "properties": {
+                    "city": {"type": "string"},
+                    "temp": {"type": "number"},
+                },
+            },
+        }
+    }
+
+    output_config = client._prepare_output_config(dict_schema)
+
+    assert output_config is not None
+    json_schema = output_config["textFormat"]["structure"]["jsonSchema"]
+    assert json_schema["name"] == "weather_output"
+    schema_parsed = json.loads(json_schema["schema"])
+    assert schema_parsed["type"] == "object"
+    assert "city" in schema_parsed["properties"]
+
+
+def test_prepare_output_config_none_returns_none() -> None:
+    """_prepare_output_config(None) must return None."""
+    client, _ = _make_client()
+
+    result = client._prepare_output_config(None)
+
+    assert result is None
+
+
+async def test_chat_response_value_populated_streaming() -> None:
+    """In streaming mode, .value should also be populated on the final response."""
+    json_response = json.dumps({"city": "Portland", "temperature": 68.0, "summary": "Cloudy"})
+    client, stub = _make_client(response_text=json_response)
+    messages = _user_messages()
+
+    stream = client.get_response(
+        messages=messages,
+        stream=True,
+        options={"max_tokens": 100, "response_format": WeatherReport},
+    )
+
+    # Consume stream and get final response
+    async for _ in stream:
+        pass
+    response = await stream.get_final_response()
+
+    assert response.value is not None
+    assert isinstance(response.value, WeatherReport)
+    assert response.value.city == "Portland"
+
+    # Verify outputConfig was sent
+    assert len(stub.calls) == 1
+    assert "outputConfig" in stub.calls[0]
+
+
+async def test_unsupported_model_validation_exception() -> None:
+    """When a model doesn't support outputConfig, a clear error should be raised."""
+    class _FailingStubBedrockRuntime:
+        def converse(self, **kwargs: Any) -> dict[str, Any]:
+            # Simulate botocore ClientError for ValidationException
+            error_response = {"Error": {"Code": "ValidationException", "Message": "Invalid field outputConfig"}}
+            raise ClientError(error_response, "Converse")
+
+    client = BedrockChatClient(
+        model="us.anthropic.claude-v2",
+        region="us-east-1",
+        client=_FailingStubBedrockRuntime(),
+    )
+
+    with pytest.raises(ValueError) as exc:
+        await client.get_response(
+            messages=_user_messages(),
+            options={"response_format": WeatherReport},
+        )
+
+    assert "does not support structured output via outputConfig.textFormat" in str(exc.value)
+    assert "Check the model's Bedrock Converse outputConfig/textFormat support." in str(exc.value)
+
+
+def test_invalid_response_format_type_raises() -> None:
+    """Non-dict, non-BaseModel response_format should raise TypeError."""
+    client, _ = _make_client()
+    with pytest.raises(TypeError, match="Pydantic BaseModel subclass"):
+        client._prepare_output_config("not_a_valid_format")
+
+
+def test_mapping_response_format_accepted() -> None:
+    """A non-dict Mapping response_format must be accepted and produce
+    correct outputConfig, not raise TypeError."""
+    from collections.abc import MutableMapping
+
+    class _WrappedMapping(MutableMapping):
+        def __init__(self, data):
+            self._data = dict(data)
+
+        def __getitem__(self, key):
+            return self._data[key]
+
+        def __setitem__(self, key, value):
+            self._data[key] = value
+
+        def __delitem__(self, key):
+            del self._data[key]
+
+        def __iter__(self):
+            return iter(self._data)
+
+        def __len__(self):
+            return len(self._data)
+
+    client, _ = _make_client()
+    mapping_format = _WrappedMapping({
+        "json_schema": {
+            "name": "test_output",
+            "schema": {
+                "type": "object",
+                "properties": {"result": {"type": "string"}},
+            },
+        }
+    })
+
+    output_config = client._prepare_output_config(mapping_format)
+
+    assert output_config is not None
+    json_schema = output_config["textFormat"]["structure"]["jsonSchema"]
+    assert json_schema["name"] == "test_output"
+    schema = json.loads(json_schema["schema"])
+    assert schema.get("additionalProperties") is False
+
+
+def test_shape_b_dict_schema_wire_format() -> None:
+    """Dict response_format in Shape B (inner shape directly) should
+    produce correct outputConfig."""
+    client, _ = _make_client()
+
+    response_format = {
+        "name": "weather_output",
+        "schema": {
+            "type": "object",
+            "properties": {
+                "city": {"type": "string"},
+                "temperature": {"type": "number"},
+            },
+        },
+    }
+
+    output_config = client._prepare_output_config(response_format)
+
+    assert output_config is not None
+    text_format = output_config["textFormat"]
+    assert text_format["type"] == "json_schema"
+    json_schema = text_format["structure"]["jsonSchema"]
+    assert json_schema["name"] == "weather_output"
+    schema = json.loads(json_schema["schema"])
+    assert schema.get("additionalProperties") is False
+
+
+def test_dict_schema_not_mutated() -> None:
+    """Caller's dict schema must not be mutated by _prepare_output_config."""
+    client, _ = _make_client()
+    original_schema = {
+        "json_schema": {
+            "name": "test",
+            "schema": {
+                "type": "object",
+                "properties": {"a": {"type": "string"}},
+            },
+        }
+    }
+    snapshot = copy.deepcopy(original_schema)
+    client._prepare_output_config(original_schema)
+    assert original_schema == snapshot, "Original dict schema was mutated"
+
+
+async def test_non_outputconfig_validation_exception_propagates() -> None:
+    """ValidationException unrelated to outputConfig must propagate
+    as raw ClientError, not be caught and reclassified."""
+    client, _ = _make_client()
+    error_response = {
+        "Error": {
+            "Code": "ValidationException",
+            "Message": "Invalid message format",
+        }
+    }
+    with (
+        patch.object(
+            client,
+            "_bedrock_client",
+            **{"converse.side_effect": ClientError(error_response, "Converse")},
+        ),
+        pytest.raises(ClientError),
+    ):
+        await client.get_response(
+            messages=_user_messages(),
+            options={"max_tokens": 100},
+        )
+
+
+# endregion

python/packages/core/AGENTS.md

@@ -56,7 +56,7 @@ agent_framework/
 - **`AgentMiddleware`** - Intercepts agent `run()` calls
 - **`ChatMiddleware`** - Intercepts chat client `get_response()` calls
 - **`FunctionMiddleware`** - Intercepts function/tool invocations
-- **`AgentContext`** / **`ChatContext`** / **`FunctionInvocationContext`** - Context objects passed through middleware
+- **`AgentContext`** / **`ChatContext`** / **`FunctionInvocationContext`** - Context objects passed through middleware. A tool can declare a `FunctionInvocationContext` parameter to receive it; `context.tools` is the live, mutable tools list for the run, and `context.add_tools(...)` / `context.remove_tools(...)` enable progressive tool exposure (changes apply on the next function-calling iteration).
 
 ### Sessions (`_sessions.py`)
 
@@ -76,6 +76,19 @@ agent_framework/
 - **`SkillScriptRunner`** - Protocol for file-based script execution. Any callable matching `(skill, script, args) -> Any` satisfies it. Code-defined scripts do not use a runner.
 - **`SkillsProvider`** - Context provider (extends `ContextProvider`) that discovers file-based skills from `SKILL.md` files and/or accepts code-defined `Skill` instances. Follows progressive disclosure: advertise → load → read resources / run scripts.
 
+### Model Context Protocol (`_mcp.py`)
+
+- **`MCPTool`** - Base wrapper that owns the MCP `ClientSession` and exposes the remote server's tools as `FunctionTool`s.
+- **`MCPStdioTool`** / **`MCPStreamableHTTPTool`** / **`MCPWebsocketTool`** - Transport-specific subclasses.
+- **`MCPTaskOptions`** (experimental, `MCP_LONG_RUNNING_TASKS` feature, **frozen**) - Per-tool-instance options controlling the SEP-2663 long-running task lifecycle. When the server advertises a tool with `execution.taskSupport == "required"`, `MCPTool.call_tool` transparently routes through `call_tool_as_task`, which sends an augmented `tools/call`, polls `tasks/get` until terminal, and reinterprets `tasks/result` as a normal `CallToolResult`. Instances are immutable; replace via `MCPTool.task_options = MCPTaskOptions(...)`. Fields:
+  - `default_ttl: timedelta | None` — forwarded to the server as `params.task.ttl` (milliseconds). When `None`, the server's default applies.
+  - `cancel_remote_task_on_local_cancellation: bool = True` — only gates the `CancelledError` path. Abandonment paths (see below) always cancel.
+  - `max_task_wait: timedelta | None` — client-side deadline for the whole post-create lifecycle (poll + result fetch). When exceeded, raises `ToolExecutionException` and fires a best-effort `tasks/cancel`. `None` (default) means no client-side bound. Bounds sleeps, sends, AND reconnects via `asyncio.wait_for`.
+- **Permissive fallback**: servers that ignore the augmentation (return `CallToolResult` directly) or reject the unknown `task` field with `METHOD_NOT_FOUND` / `INVALID_PARAMS` fall back to the plain `session.call_tool(...)` path so legacy servers keep working. An unparseable success response (server accepted the augmented call but returned a payload that is neither `CreateTaskResult` nor `CallToolResult`) **does not** fall back — it raises `ToolExecutionException` to avoid double-executing a side-effecting tool.
+- **Submit-vs-track reconnect policy**: a dropped connection before a `task_id` is known raises `ToolExecutionException("connection lost; task state unknown")` without re-issuing the augmented `tools/call`, so a server that accepted the request but lost the response cannot be made to start the same operation twice; once a `task_id` exists, `tasks/get` / `tasks/result` reconnect once and retry against the same id (a shared `_send_with_one_reconnect` helper).
+- **Cancel-on-abandonment vs terminal failure**: any path where the remote task may still be running (max-wait exceeded, hard `McpError` in poll, malformed `tasks/get`, second connection loss in poll/fetch, reconnect failure) fires best-effort `tasks/cancel` before raising. Terminal failures (`failed`/`cancelled`/`input_required` server-side, `completed+isError`, malformed `tasks/result` after server completed) do **not** cancel — the server is already done. `_MCPTaskAbandoned` is the private marker distinguishing the two.
+- **Transient poll retry**: a slow `tasks/get` that surfaces as `McpError(code=408 REQUEST_TIMEOUT)` is retried (bounded by `max_task_wait`). All other non-connection `McpError`s during poll are treated as abandonment. `tasks/result` does not get transient retry — the server has already completed, so a slow payload fetch is anomalous.
+
 ### File Access Harness (`_harness/_file_access.py`)
 
 - **`AgentFileStore`** - Abstract async store backing the file-access harness. Implementations expose `write_file`, `read_file`, `delete_file`, `list_files`, `file_exists`, `search_files`, and `create_directory` over forward-slash relative paths.

python/packages/core/agent_framework/__init__.py

@@ -71,6 +71,7 @@ from ._evaluation import (
     Evaluator,
     ExpectedToolCall,
     LocalEvaluator,
+    RubricScore,
     evaluate_agent,
     evaluate_workflow,
     evaluator,
@@ -123,7 +124,7 @@ from ._harness._todo import (
     TodoSessionStore,
     TodoStore,
 )
-from ._mcp import MCPStdioTool, MCPStreamableHTTPTool, MCPWebsocketTool
+from ._mcp import MCPStdioTool, MCPStreamableHTTPTool, MCPTaskOptions, MCPWebsocketTool
 from ._middleware import (
     AgentContext,
     AgentMiddleware,
@@ -167,6 +168,9 @@ from ._skills import (
     InlineSkillResource,
     InlineSkillScript,
     InMemorySkillsSource,
+    MCPSkill,
+    MCPSkillResource,
+    MCPSkillsSource,
     Skill,
     SkillFrontmatter,
     SkillResource,
@@ -440,8 +444,12 @@ __all__ = [
     "InlineSkillResource",
     "InlineSkillScript",
     "LocalEvaluator",
+    "MCPSkill",
+    "MCPSkillResource",
+    "MCPSkillsSource",
     "MCPStdioTool",
     "MCPStreamableHTTPTool",
+    "MCPTaskOptions",
     "MCPWebsocketTool",
     "MemoryContextProvider",
     "MemoryFileStore",
@@ -460,6 +468,7 @@ __all__ = [
     "ResponseStream",
     "Role",
     "RoleLiteral",
+    "RubricScore",
     "RunContext",
     "Runner",
     "RunnerContext",

python/packages/core/agent_framework/_evaluation.py

@@ -311,12 +311,15 @@ class EvalScoreResult:
         score: Numeric score from the evaluator.
         passed: Whether the item passed this evaluator's threshold.
         sample: Optional raw evaluator output (rationale, metadata).
+        dimensions: Per-dimension scores when this evaluator is a rubric
+            evaluator.  ``None`` for non-rubric (e.g. built-in) evaluators.
     """
 
     name: str
     score: float
     passed: bool | None = None
     sample: dict[str, Any] | None = None
+    dimensions: list[RubricScore] | None = None
 
 
 @experimental(feature_id=ExperimentalFeature.EVALS)
@@ -496,6 +499,179 @@ class EvalResults:
                     detail += f" Errored items: {', '.join(summaries)}."
             raise EvalNotPassedError(detail)
 
+    def assert_score_at_least(
+        self,
+        min_score: float,
+        *,
+        evaluator: str | None = None,
+        msg: str | None = None,
+    ) -> None:
+        """Assert every item's score (optionally filtered by evaluator) is ``>= min_score``.
+
+        Designed for CI gates on generated rubric evaluators (e.g.
+        ``results.assert_score_at_least(0.80)``).  Includes any
+        sub-results from workflow evaluations.
+
+        Args:
+            min_score: Minimum acceptable score (inclusive).
+            evaluator: When set, only check scores from the evaluator
+                whose ``EvalScoreResult.name`` matches.
+            msg: Optional custom failure message.
+
+        Raises:
+            EvalNotPassedError: When any matching score is below the threshold.
+        """
+        offenders: list[str] = []
+
+        def _check(results: EvalResults) -> None:
+            for item in results.items:
+                for score in item.scores:
+                    if evaluator is not None and score.name != evaluator:
+                        continue
+                    if score.score < min_score:
+                        offenders.append(f"{item.item_id}/{score.name}={score.score:.3f}")
+            for sub in results.sub_results.values():
+                _check(sub)
+
+        _check(self)
+        if offenders:
+            detail = msg or (
+                f"{len(offenders)} score(s) below threshold {min_score}"
+                f"{' for ' + evaluator if evaluator else ''}: {', '.join(offenders[:5])}"
+                + (f" (+{len(offenders) - 5} more)" if len(offenders) > 5 else "")
+            )
+            raise EvalNotPassedError(detail)
+
+    def assert_dimension_score_at_least(
+        self,
+        dimension_id: str,
+        min_score: float,
+        *,
+        evaluator: str | None = None,
+        require_applicable: bool = False,
+        msg: str | None = None,
+    ) -> None:
+        """Assert every item's score for a rubric *dimension* is ``>= min_score``.
+
+        Walks ``EvalScoreResult.dimensions`` looking for the named
+        dimension across all items (and sub-results).  Non-applicable
+        dimensions are skipped by default; pass
+        ``require_applicable=True`` to fail when no applicable score is
+        produced.
+
+        Args:
+            dimension_id: Dimension id (matches the rubric definition).
+            min_score: Minimum acceptable dimension score (inclusive).
+            evaluator: When set, only consider scores from the evaluator
+                whose ``EvalScoreResult.name`` matches.
+            require_applicable: When ``True``, missing or non-applicable
+                dimension scores raise.  Defaults to ``False`` (skip).
+            msg: Optional custom failure message.
+
+        Raises:
+            EvalNotPassedError: When the dimension fails the threshold.
+        """
+        offenders: list[str] = []
+        missing_items: list[str] = []
+
+        def _check(results: EvalResults) -> None:
+            for item in results.items:
+                found_applicable = False
+                for score in item.scores:
+                    if evaluator is not None and score.name != evaluator:
+                        continue
+                    if not score.dimensions:
+                        continue
+                    for rs in score.dimensions:
+                        if rs.id != dimension_id:
+                            continue
+                        if not rs.applicable:
+                            continue
+                        found_applicable = True
+                        if rs.score is None or rs.score < min_score:
+                            offenders.append(
+                                f"{item.item_id}/{score.name}/{dimension_id}="
+                                f"{rs.score if rs.score is not None else 'None'}"
+                            )
+                if require_applicable and not found_applicable:
+                    missing_items.append(item.item_id)
+            for sub in results.sub_results.values():
+                _check(sub)
+
+        _check(self)
+        problems: list[str] = []
+        if offenders:
+            problems.append(
+                f"{len(offenders)} dimension score(s) for '{dimension_id}' below {min_score}: "
+                f"{', '.join(offenders[:5])}" + (f" (+{len(offenders) - 5} more)" if len(offenders) > 5 else "")
+            )
+        if missing_items:
+            problems.append(
+                f"Dimension '{dimension_id}' not applicable on {len(missing_items)} item(s): "
+                f"{', '.join(missing_items[:5])}"
+            )
+        if problems:
+            raise EvalNotPassedError(msg or "; ".join(problems))
+
+    def assert_no_failed_items(self, msg: str | None = None) -> None:
+        """Assert no item ended in ``fail`` or ``error`` status.
+
+        Includes any sub-results from workflow evaluations.
+
+        Args:
+            msg: Optional custom failure message.
+
+        Raises:
+            EvalNotPassedError: When any item failed or errored.
+        """
+        bad: list[str] = []
+
+        def _check(results: EvalResults) -> None:
+            for item in results.items:
+                if item.is_failed or item.is_error:
+                    bad.append(f"{item.item_id}:{item.status}")
+            for sub in results.sub_results.values():
+                _check(sub)
+
+        _check(self)
+        if bad:
+            detail = msg or (
+                f"{len(bad)} item(s) failed or errored: {', '.join(bad[:5])}"
+                + (f" (+{len(bad) - 5} more)" if len(bad) > 5 else "")
+            )
+            raise EvalNotPassedError(detail)
+
+
+# endregion
+
+# region Generated rubric evaluators
+
+
+@experimental(feature_id=ExperimentalFeature.EVALS)
+@dataclass(frozen=True)
+class RubricScore:
+    """A single dimension's score from a rubric-based evaluator run.
+
+    Rubric evaluators emit one ``RubricScore`` per dimension per item.
+    Attached to :class:`EvalScoreResult` as a typed view of the raw
+    ``properties.rubric_scores`` payload returned by providers such as
+    Foundry's generated rubric evaluators.
+
+    Attributes:
+        id: Dimension id (matches the rubric definition).
+        score: Numeric score, or ``None`` when the dimension was marked
+            non-applicable for this item.
+        applicable: Whether the dimension applied to this item.
+        weight: Dimension weight (mirrors the rubric definition).
+        reason: Short rationale produced by the evaluator.
+    """
+
+    id: str
+    score: int | None
+    applicable: bool
+    weight: int
+    reason: str
+
 
 # endregion
 

python/packages/core/agent_framework/_feature_stage.py

@@ -50,6 +50,7 @@ class ExperimentalFeature(str, Enum):
     on enum membership or attribute presence over time.
     """
 
+    DECLARATIVE_AGENTS = "DECLARATIVE_AGENTS"
     EVALS = "EVALS"
     FILE_HISTORY = "FILE_HISTORY"
     FIDES = "FIDES"
@@ -57,6 +58,9 @@ class ExperimentalFeature(str, Enum):
     FOUNDRY_PREVIEW_TOOLS = "FOUNDRY_PREVIEW_TOOLS"
     FUNCTIONAL_WORKFLOWS = "FUNCTIONAL_WORKFLOWS"
     HARNESS = "HARNESS"
+    MCP_LONG_RUNNING_TASKS = "MCP_LONG_RUNNING_TASKS"
+    MCP_SKILLS = "MCP_SKILLS"
+    PROGRESSIVE_TOOLS = "PROGRESSIVE_TOOLS"
     SKILLS = "SKILLS"
     TO_PROMPT_AGENT = "TO_PROMPT_AGENT"
 

python/packages/core/agent_framework/_harness/_agent.py

@@ -14,12 +14,13 @@ import logging
 from collections.abc import Callable, Sequence
 from typing import TYPE_CHECKING, Any
 
-from .._agents import Agent
+from .._agents import Agent, SupportsAgentRun
 from .._clients import SupportsWebSearchTool
 from .._compaction import CompactionProvider, ContextWindowCompactionStrategy, ToolResultCompactionStrategy
 from .._feature_stage import ExperimentalFeature, experimental
 from .._sessions import ContextProvider, HistoryProvider, InMemoryHistoryProvider
 from .._skills import SkillsProvider
+from ._background_agents import BackgroundAgentsProvider
 from ._memory import MemoryContextProvider, MemoryStore
 from ._mode import AgentModeProvider
 from ._todo import TodoProvider
@@ -103,6 +104,8 @@ def _assemble_context_providers(
     memory_store: MemoryStore | None,
     skills_provider: SkillsProvider | None,
     skills_paths: Sequence[str] | None,
+    background_agents: Sequence[SupportsAgentRun] | None,
+    background_agents_instructions: str | None,
     extra_context_providers: Sequence[ContextProvider] | None,
 ) -> list[ContextProvider]:
     """Assemble the ordered list of context providers."""
@@ -130,6 +133,10 @@ def _assemble_context_providers(
     if skills_paths:
         providers.append(SkillsProvider.from_paths(*skills_paths))
 
+    # Background agents are opt-in: only added when agents are provided.
+    if background_agents:
+        providers.append(BackgroundAgentsProvider(background_agents, instructions=background_agents_instructions))
+
     # Append any user-supplied additional providers.
     if extra_context_providers:
         providers.extend(extra_context_providers)
@@ -165,6 +172,8 @@ def create_harness_agent(
     memory_store: MemoryStore | None = None,
     skills_provider: SkillsProvider | None = None,
     skills_paths: Sequence[str] | None = None,
+    background_agents: Sequence[SupportsAgentRun] | None = None,
+    background_agents_instructions: str | None = None,
     disable_web_search: bool = False,
     otel_provider_name: str | None = None,
     context_providers: Sequence[ContextProvider] | None = None,
@@ -182,6 +191,7 @@ def create_harness_agent(
     - **AgentModeProvider** — plan/execute mode tracking
     - **MemoryContextProvider** — file-based durable memory (when ``memory_store`` provided)
     - **SkillsProvider** — skill discovery and progressive loading
+    - **BackgroundAgentsProvider** — delegate work to background sub-agents
     - **OpenTelemetry** — observability via ``AgentTelemetryLayer``
 
     Each feature can be disabled or customized via keyword arguments.
@@ -253,6 +263,13 @@ def create_harness_agent(
         skills_paths: Paths for file-based skill discovery (looks for SKILL.md files).
             Can be combined with ``skills_provider``. When neither ``skills_provider``
             nor ``skills_paths`` is provided, no SkillsProvider is added.
+        background_agents: Collection of agents available for background task delegation.
+            When provided, a ``BackgroundAgentsProvider`` is automatically included,
+            enabling the agent to start, monitor, and retrieve results from background tasks.
+            Each agent must have a non-empty, unique name (case-insensitive).
+        background_agents_instructions: Optional instruction override for the
+            ``BackgroundAgentsProvider``. May include ``{background_agents}`` placeholder
+            which will be replaced with the agent listing.
         disable_web_search: When True, skip automatic web search tool inclusion.
             When False (default), the web search tool is automatically added if the
             client implements SupportsWebSearchTool. A warning is logged if the client
@@ -302,6 +319,8 @@ def create_harness_agent(
         memory_store=memory_store,
         skills_provider=skills_provider,
         skills_paths=skills_paths,
+        background_agents=background_agents,
+        background_agents_instructions=background_agents_instructions,
         extra_context_providers=context_providers,
     )
 

python/packages/core/agent_framework/_mcp.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import asyncio
 import base64
+import contextlib
 import contextvars
 import json
 import logging
@@ -12,12 +13,14 @@ import sys
 from abc import abstractmethod
 from collections.abc import Callable, Collection, Coroutine, Mapping, Sequence
 from contextlib import AsyncExitStack, _AsyncGeneratorContextManager  # type: ignore
+from dataclasses import dataclass
 from datetime import timedelta
 from functools import partial
 from typing import TYPE_CHECKING, Any, Literal, TypedDict, cast
 
 from opentelemetry import propagate
 
+from ._feature_stage import ExperimentalFeature, experimental
 from ._tools import FunctionTool
 from ._types import (
     ChatOptions,
@@ -149,6 +152,73 @@ def _url_origin(url: Any) -> tuple[str, str, int | None]:
     return (url.scheme, url.host or "", port)
 
 
+# Internal polling bounds for MCP long-running tasks. Not user-tunable today;
+# promote to MCPTaskOptions if a concrete need arises.
+_MCP_TASK_MIN_POLL_INTERVAL = timedelta(milliseconds=500)
+_MCP_TASK_MAX_POLL_INTERVAL = timedelta(seconds=5)
+_MCP_TASK_CANCEL_TIMEOUT = timedelta(seconds=5)
+_MCP_TASK_TERMINAL_STATUSES: frozenset[str] = frozenset({"completed", "failed", "cancelled", "input_required"})
+
+# Total send attempts for a Phase 2 request (initial try + one reconnect-and-retry).
+# A single transient disconnect should not abort a long-running task; sustained outages
+# surface as ``_MCPTaskAbandoned`` after the second failure.
+_MCP_RECONNECT_ATTEMPTS = 2
+
+
+class _MCPTaskAbandoned(ToolExecutionException):
+    """Raised when the remote MCP task may still be running and must be cancelled.
+
+    Subclass of ToolExecutionException so callers see a normal tool failure.
+    """
+
+
+class _MCPDeadlineExpired(Exception):
+    """Internal marker for ``max_task_wait`` expiry; distinct from inner TimeoutError."""
+
+
+@experimental(feature_id=ExperimentalFeature.MCP_LONG_RUNNING_TASKS)
+@dataclass(frozen=True)
+class MCPTaskOptions:
+    """Options controlling how MCPTool drives the MCP long-running task lifecycle.
+
+    When an MCP server advertises a tool with ``execution.taskSupport == "required"``,
+    the framework transparently drives the SEP-2663 ``tools/call`` → ``tasks/get``
+    (polled) → ``tasks/result`` lifecycle so the agent sees a normal tool result.
+
+    Instances are immutable; replace the whole object via
+    ``MCPTool.task_options = MCPTaskOptions(...)`` to change behavior.
+
+    Attributes:
+        default_ttl: Optional task-record retention time forwarded to the server as
+            ``params.task.ttl`` (milliseconds, integer). The server keeps the task
+            record around this long after the task reaches a terminal status so the
+            client can still call ``tasks/get`` / ``tasks/result``; it does not
+            cancel a running task. When ``None``, the server applies its own default.
+            Must be positive if set (zero would expire the record before any client
+            could read it).
+        cancel_remote_task_on_local_cancellation: If True (default), a local
+            cancellation of the awaiting coroutine triggers a best-effort
+            ``tasks/cancel`` on the server before re-raising ``CancelledError``.
+            Only gates ``CancelledError``; abandonment paths (max-wait,
+            unrecoverable poll errors, lost connection after task_id is known)
+            always cancel regardless of this flag.
+        max_task_wait: Optional client-side deadline for the whole post-create
+            lifecycle (poll + result fetch). When exceeded, raises
+            ``ToolExecutionException`` and fires a best-effort ``tasks/cancel``.
+            ``None`` (default) means no client-side bound. Must be positive if set.
+    """
+
+    default_ttl: timedelta | None = None
+    cancel_remote_task_on_local_cancellation: bool = True
+    max_task_wait: timedelta | None = None
+
+    def __post_init__(self) -> None:
+        if self.default_ttl is not None and self.default_ttl.total_seconds() <= 0:
+            raise ValueError("MCPTaskOptions.default_ttl must be positive.")
+        if self.max_task_wait is not None and self.max_task_wait.total_seconds() <= 0:
+            raise ValueError("MCPTaskOptions.max_task_wait must be positive.")
+
+
 def streamable_http_client(*args: Any, **kwargs: Any) -> _AsyncGeneratorContextManager[Any, None]:
     """Lazily import the MCP streamable HTTP transport."""
     try:
@@ -217,6 +287,7 @@ class MCPTool:
         request_timeout: int | None = None,
         client: SupportsChatGetResponse | None = None,
         additional_properties: dict[str, Any] | None = None,
+        task_options: MCPTaskOptions | None = None,
     ) -> None:
         """Initialize the MCP Tool base.
 
@@ -248,6 +319,9 @@ class MCPTool:
             request_timeout: Timeout in seconds for MCP requests.
             client: A chat client for sampling callbacks.
             additional_properties: Additional properties for the tool.
+            task_options: Options controlling how long-running MCP tasks are driven for
+                tools that advertise ``execution.taskSupport == "required"``. When ``None``,
+                the defaults from :class:`MCPTaskOptions` are used.
         """
         self.name = name
         self.description = description or ""
@@ -259,6 +333,10 @@ class MCPTool:
         self.parse_tool_results = parse_tool_results
         self.load_prompts_flag = load_prompts
         self.parse_prompt_results = parse_prompt_results
+        # Defer constructing the default MCPTaskOptions so the experimental warning
+        # only fires when LRO is actually engaged (lazy-resolved by _effective_task_options).
+        self._task_options_explicit: MCPTaskOptions | None = task_options
+        self._task_options_default: MCPTaskOptions | None = None
         self._exit_stack = AsyncExitStack()
         self._lifecycle_lock = asyncio.Lock()
         self._lifecycle_request_lock = asyncio.Lock()
@@ -270,6 +348,7 @@ class MCPTool:
         self.client = client
         self._functions: list[FunctionTool] = []
         self._tool_call_meta_by_name: dict[str, dict[str, Any]] = {}
+        self._tool_task_support_by_name: dict[str, str] = {}
         self.is_connected: bool = False
         self._tools_loaded: bool = False
         self._prompts_loaded: bool = False
@@ -1131,6 +1210,7 @@ class MCPTool:
         # Track existing function names to prevent duplicates
         existing_names = {func.name for func in self._functions}
         tool_call_meta_by_name: dict[str, dict[str, Any]] = {}
+        tool_task_support_by_name: dict[str, str] = {}
 
         params: types.PaginatedRequestParams | None = None
         while True:
@@ -1168,6 +1248,10 @@ class MCPTool:
                 if tool.meta is not None:
                     tool_call_meta_by_name[tool.name] = dict(tool.meta)
 
+                task_support = getattr(getattr(tool, "execution", None), "taskSupport", None)
+                if task_support is not None:
+                    tool_task_support_by_name[tool.name] = task_support
+
                 normalized_name = _normalize_mcp_name(tool.name)
                 local_name = _build_prefixed_mcp_name(normalized_name, self.tool_name_prefix)
 
@@ -1216,6 +1300,7 @@ class MCPTool:
             params = types.PaginatedRequestParams(cursor=tool_list.nextCursor)
 
         self._tool_call_meta_by_name = tool_call_meta_by_name
+        self._tool_task_support_by_name = tool_task_support_by_name
 
     async def _close_on_owner(self) -> None:
         # Cancel any pending reload tasks before tearing down the session.
@@ -1292,6 +1377,29 @@ class MCPTool:
                     inner_exception=ex,
                 ) from ex
 
+    def _effective_task_options(self) -> MCPTaskOptions:
+        """Return the effective MCPTaskOptions, lazily constructing defaults on first use.
+
+        Defers the implicit ``MCPTaskOptions()`` so the experimental warning only
+        fires when LRO is actually engaged (server advertises ``taskSupport=required``).
+        """
+        explicit = self._task_options_explicit
+        if explicit is not None:
+            return explicit
+        if self._task_options_default is None:
+            self._task_options_default = MCPTaskOptions()
+        return self._task_options_default
+
+    @property
+    def task_options(self) -> MCPTaskOptions:
+        """The effective MCPTaskOptions for this tool (lazy defaults)."""
+        return self._effective_task_options()
+
+    @task_options.setter
+    def task_options(self, value: MCPTaskOptions | None) -> None:
+        self._task_options_explicit = value
+        self._task_options_default = None
+
     async def call_tool(self, tool_name: str, **kwargs: Any) -> str | list[Content]:
         """Call a tool with the given arguments.
 
@@ -1322,47 +1430,12 @@ class MCPTool:
                 "Tools are not loaded for this server, please set load_tools=True in the constructor."
             )
 
-        raw_user_meta: object | None = kwargs.get("_meta")
-        user_meta: dict[str, Any] | None = None
-        if raw_user_meta is not None and not isinstance(raw_user_meta, dict):
-            raise ToolExecutionException("MCP tool metadata provided via _meta must be a dict.")
-        if isinstance(raw_user_meta, dict):
-            raw_user_meta_dict = cast(Mapping[object, object], raw_user_meta)
-            user_meta = {}
-            for key, value in raw_user_meta_dict.items():
-                if not isinstance(key, str):
-                    raise ToolExecutionException("MCP tool metadata provided via _meta must use string keys.")
-                user_meta[key] = value
+        # Tools advertising taskSupport == "required" cannot complete via plain tools/call;
+        # route through the long-running task lifecycle transparently.
+        if self._tool_task_support_by_name.get(tool_name) == "required":
+            return await self.call_tool_as_task(tool_name, **kwargs)
 
-        # Filter out framework kwargs that cannot be serialized by the MCP SDK.
-        # These are internal objects passed through the function invocation pipeline
-        # that should not be forwarded to external MCP servers.
-        # conversation_id is an internal tracking ID used by services like Azure AI.
-        # options contains metadata/store used by AG-UI for Azure AI client requirements.
-        # response_format is a Pydantic model class used for structured output (not serializable).
-        filtered_kwargs = {
-            k: v
-            for k, v in kwargs.items()
-            if k
-            not in {
-                "chat_options",
-                "tools",
-                "tool_choice",
-                "session",
-                "thread",
-                "conversation_id",
-                "options",
-                "response_format",
-                "_meta",
-            }
-        }
-
-        # Some MCP proxies require their tools/list metadata to be echoed on tools/call.
-        tool_meta = self._tool_call_meta_by_name.get(tool_name)
-        request_meta = dict(tool_meta) if tool_meta is not None else None
-        if user_meta is not None:
-            request_meta = {**(request_meta or {}), **user_meta}
-        meta = _inject_otel_into_mcp_meta(request_meta)
+        filtered_kwargs, meta = self._prepare_call_kwargs(tool_name, kwargs)
 
         parser = self.parse_tool_results or self._parse_tool_result_from_mcp
         # Try the operation, reconnecting once if the connection is closed
@@ -1411,6 +1484,479 @@ class MCPTool:
                 raise ToolExecutionException(f"Failed to call tool '{tool_name}'.", inner_exception=ex) from ex
         raise ToolExecutionException(f"Failed to call tool '{tool_name}' after retries.")
 
+    def _prepare_call_kwargs(
+        self, tool_name: str, kwargs: dict[str, Any]
+    ) -> tuple[dict[str, Any], dict[str, Any] | None]:
+        """Filter framework-only kwargs and build the merged MCP request metadata."""
+        raw_user_meta: object | None = kwargs.get("_meta")
+        user_meta: dict[str, Any] | None = None
+        if raw_user_meta is not None and not isinstance(raw_user_meta, dict):
+            raise ToolExecutionException("MCP tool metadata provided via _meta must be a dict.")
+        if isinstance(raw_user_meta, dict):
+            raw_user_meta_dict = cast(Mapping[object, object], raw_user_meta)
+            user_meta = {}
+            for key, value in raw_user_meta_dict.items():
+                if not isinstance(key, str):
+                    raise ToolExecutionException("MCP tool metadata provided via _meta must use string keys.")
+                user_meta[key] = value
+
+        # Filter out framework kwargs that cannot be serialized by the MCP SDK.
+        # These are internal objects passed through the function invocation pipeline
+        # that should not be forwarded to external MCP servers.
+        # conversation_id is an internal tracking ID used by services like Azure AI.
+        # options contains metadata/store used by AG-UI for Azure AI client requirements.
+        # response_format is a Pydantic model class used for structured output (not serializable).
+        filtered_kwargs = {
+            k: v
+            for k, v in kwargs.items()
+            if k
+            not in {
+                "chat_options",
+                "tools",
+                "tool_choice",
+                "session",
+                "thread",
+                "conversation_id",
+                "options",
+                "response_format",
+                "_meta",
+            }
+        }
+
+        # Some MCP proxies require their tools/list metadata to be echoed on tools/call.
+        tool_meta = self._tool_call_meta_by_name.get(tool_name)
+        request_meta = dict(tool_meta) if tool_meta is not None else None
+        if user_meta is not None:
+            request_meta = {**(request_meta or {}), **user_meta}
+        meta = _inject_otel_into_mcp_meta(request_meta)
+        return filtered_kwargs, meta
+
+    async def call_tool_as_task(self, tool_name: str, **kwargs: Any) -> str | list[Content]:
+        """Call an MCP tool via the long-running task lifecycle (SEP-2663).
+
+        Issues an augmented ``tools/call`` with ``params.task`` set from
+        ``self.task_options``, then polls ``tasks/get`` until the server reports a
+        terminal status. On ``completed`` the payload is fetched via ``tasks/result``,
+        validated as a ``CallToolResult`` and parsed identically to :meth:`call_tool`.
+
+        Local cancellation triggers a best-effort ``tasks/cancel`` (controlled by
+        :attr:`MCPTaskOptions.cancel_remote_task_on_local_cancellation`) before
+        ``asyncio.CancelledError`` is re-raised.
+
+        Args:
+            tool_name: The remote MCP tool name.
+
+        Keyword Args:
+            kwargs: Arguments forwarded to the tool. See :meth:`call_tool` for the
+                framework kwargs that are filtered out.
+
+        Returns:
+            A list of Content items (or a string when a custom ``parse_tool_results``
+            callback is configured).
+        """
+        from anyio import ClosedResourceError
+        from mcp.shared.exceptions import McpError
+
+        if not self.load_tools_flag:
+            raise ToolExecutionException(
+                "Tools are not loaded for this server, please set load_tools=True in the constructor."
+            )
+
+        filtered_kwargs, meta = self._prepare_call_kwargs(tool_name, kwargs)
+        parser = self.parse_tool_results or self._parse_tool_result_from_mcp
+
+        # Submit the task: issue augmented tools/call. Do NOT retry on connection loss here:
+        # the server may have accepted the request and created a task before the
+        # response was lost, so retrying could start the long-running operation twice.
+        # Reconnect-and-retry is only safe after the task_id is known.
+        try:
+            task_id, fallback_result = await self._call_tool_as_task_create(tool_name, filtered_kwargs, meta)
+        except (ClosedResourceError, McpError) as ex:
+            if not self._is_connection_lost(ex):
+                error_message = ex.error.message if isinstance(ex, McpError) else str(ex)
+                raise ToolExecutionException(error_message, inner_exception=ex) from ex
+            raise ToolExecutionException(
+                f"Failed to call tool '{tool_name}' - connection lost; task state unknown.",
+                inner_exception=ex,
+            ) from ex
+        except ToolExecutionException:
+            raise
+        except Exception as ex:
+            raise ToolExecutionException(f"Failed to call tool '{tool_name}'.", inner_exception=ex) from ex
+
+        # Server returned a CallToolResult (no task created) or fell back to plain tools/call.
+        if fallback_result is not None:
+            if fallback_result.isError:
+                parsed = parser(fallback_result)
+                text = (
+                    "\n".join(c.text for c in parsed if c.type == "text" and c.text)
+                    if isinstance(parsed, list)
+                    else str(parsed)
+                )
+                raise ToolExecutionException(text or str(parsed))
+            return parser(fallback_result)
+
+        if task_id is None:
+            raise ToolExecutionException(
+                f"MCP server did not return a task_id or fallback result for '{tool_name}'."
+            )
+
+        # Track to completion: poll until terminal, then fetch payload. Never re-issue
+        # tools/call past this point; reconnect-and-retry only against the same task_id.
+        opts = self._effective_task_options()
+        max_wait_s = opts.max_task_wait.total_seconds() if opts.max_task_wait is not None else None
+
+        async def _await_task_completion() -> str | list[Content]:
+            terminal = await self._poll_task_until_terminal(task_id)
+            return await self._handle_terminal_task(tool_name, task_id, terminal, parser)
+
+        try:
+            if max_wait_s is not None:
+                try:
+                    result = await self._await_with_deadline(_await_task_completion(), max_wait_s)
+                    return cast("str | list[Content]", result)
+                except _MCPDeadlineExpired as ex:
+                    self._spawn_best_effort_cancel(task_id)
+                    raise ToolExecutionException(
+                        f"MCP task '{task_id}' exceeded max_task_wait of {max_wait_s}s.",
+                        inner_exception=ex,
+                    ) from ex
+            else:
+                return await _await_task_completion()
+        except asyncio.CancelledError:
+            if opts.cancel_remote_task_on_local_cancellation:
+                self._spawn_best_effort_cancel(task_id)
+            raise
+        except _MCPTaskAbandoned:
+            # Pre-terminal abandonment (hard poll error, malformed get, second
+            # disconnect, reconnect failure): cancel + re-raise as plain
+            # ToolExecutionException to the function-calling loop.
+            self._spawn_best_effort_cancel(task_id)
+            raise
+        # Plain ToolExecutionException from terminal failures (failed/cancelled/
+        # input_required, completed+isError, malformed result post-completion)
+        # propagates without cancel — server is already done.
+
+    async def _call_tool_as_task_create(
+        self, tool_name: str, arguments: dict[str, Any], meta: dict[str, Any] | None
+    ) -> tuple[str | None, types.CallToolResult | None]:
+        """Send the augmented tools/call.
+
+        Returns ``(task_id, None)`` when the server created a task,
+        ``(None, CallToolResult)`` when it returned a non-task result, falling back
+        to plain ``tools/call`` if the server rejects the ``task`` field outright.
+        """
+        from mcp import types
+        from mcp.shared.exceptions import McpError
+        from pydantic import ValidationError
+
+        opts = self._effective_task_options()
+        ttl_ms: int | None = None
+        if opts.default_ttl is not None:
+            ttl_ms = int(opts.default_ttl.total_seconds() * 1000)
+        # Always send TaskMetadata to mark the call as task-augmented; ttl may be omitted.
+        task_metadata = types.TaskMetadata(ttl=ttl_ms)
+
+        request_meta = types.RequestParams.Meta(**meta) if meta else None
+        params = types.CallToolRequestParams(
+            name=tool_name,
+            arguments=arguments,
+            task=task_metadata,
+            _meta=request_meta,  # type: ignore[call-arg]
+        )
+        request = types.ClientRequest(types.CallToolRequest(params=params))
+
+        # Use the lenient Result type so we can extract the task_id even when
+        # the strict CreateTaskResult schema rejects the payload (the MCP Python
+        # SDK requires Task.ttl, but servers may legitimately omit it).
+        try:
+            lenient = await self.session.send_request(  # type: ignore[union-attr]
+                request,
+                types.Result,
+            )
+        except McpError as ex:
+            if ex.error.code not in (types.METHOD_NOT_FOUND, types.INVALID_PARAMS):
+                raise
+            logger.debug(
+                "Server rejected augmented tools/call for '%s' (code=%s); falling back.",
+                tool_name,
+                ex.error.code,
+            )
+            fallback = await self.session.call_tool(tool_name, arguments=arguments, meta=meta)  # type: ignore[union-attr]
+            return None, fallback
+
+        # Inspect the raw payload: a CreateTaskResult carries `task.taskId`;
+        # a legacy CallToolResult carries `content` and/or `isError`.
+        raw: dict[str, Any] = lenient.model_dump(by_alias=True, exclude_none=True)
+        raw.pop("_meta", None)
+
+        task_field = raw.get("task")
+        if isinstance(task_field, dict):
+            task_id_val = cast(dict[str, Any], task_field).get("taskId")
+            if isinstance(task_id_val, str):
+                return task_id_val, None
+
+        try:
+            legacy = types.CallToolResult.model_validate(raw)
+        except ValidationError as ex:
+            # Augmented call succeeded server-side; re-issuing a plain tools/call
+            # could double-execute a side-effecting tool.
+            raise ToolExecutionException(
+                f"MCP server returned an unparseable response to augmented tools/call "
+                f"for '{tool_name}'; cannot safely retry (server may have started the operation).",
+                inner_exception=ex,
+            ) from ex
+
+        return None, legacy
+
+    async def _poll_task_until_terminal(self, task_id: str) -> types.GetTaskResult:
+        """Poll ``tasks/get`` until the task reaches a terminal status."""
+        import httpx
+        from mcp import types
+        from mcp.shared.exceptions import McpError
+
+        # SDK raises McpError(code=httpx.REQUEST_TIMEOUT=408) on session read timeout.
+        transient_codes: frozenset[int] = frozenset({int(httpx.codes.REQUEST_TIMEOUT)})
+
+        while True:
+            request = types.ClientRequest(
+                types.GetTaskRequest(params=types.GetTaskRequestParams(taskId=task_id))
+            )
+            try:
+                # GetTaskResult.ttl is required-but-Optional in the SDK; coerce below.
+                lenient = await self._send_with_one_reconnect(
+                    request, types.Result, operation="tasks/get", task_id=task_id
+                )
+            except McpError as ex:
+                if ex.error.code in transient_codes:
+                    logger.debug(
+                        "Transient %s on tasks/get for '%s'; will retry.", ex.error.code, task_id
+                    )
+                    await asyncio.sleep(_MCP_TASK_MIN_POLL_INTERVAL.total_seconds())
+                    continue
+                # Hard server error mid-poll: task may still be running.
+                raise _MCPTaskAbandoned(ex.error.message, inner_exception=ex) from ex
+
+            try:
+                snapshot = self._coerce_get_task_result(lenient, task_id)
+            except ToolExecutionException as ex:
+                # Malformed tasks/get response; task may still be running.
+                raise _MCPTaskAbandoned(str(ex), inner_exception=ex) from ex
+
+            if snapshot.status in _MCP_TASK_TERMINAL_STATUSES:
+                return snapshot
+
+            await asyncio.sleep(self._compute_poll_delay(snapshot.pollInterval).total_seconds())
+
+    @staticmethod
+    def _coerce_get_task_result(lenient: types.Result, task_id: str) -> types.GetTaskResult:
+        """Coerce a lenient Result into GetTaskResult, defaulting ``ttl`` when absent."""
+        from mcp import types
+
+        raw = lenient.model_dump(by_alias=True, exclude_none=True)
+        raw.pop("_meta", None)
+        raw.setdefault("ttl", None)
+        try:
+            return types.GetTaskResult.model_validate(raw)
+        except Exception as ex:
+            raise ToolExecutionException(
+                f"MCP server returned a malformed tasks/get response for task '{task_id}'.",
+                inner_exception=ex,
+            ) from ex
+
+    @staticmethod
+    def _compute_poll_delay(server_interval_ms: int | None) -> timedelta:
+        """Clamp the server-suggested poll interval to ``[min, max]``."""
+        if server_interval_ms is None or server_interval_ms <= 0:
+            return _MCP_TASK_MIN_POLL_INTERVAL
+        suggested = timedelta(milliseconds=server_interval_ms)
+        if suggested < _MCP_TASK_MIN_POLL_INTERVAL:
+            return _MCP_TASK_MIN_POLL_INTERVAL
+        if suggested > _MCP_TASK_MAX_POLL_INTERVAL:
+            return _MCP_TASK_MAX_POLL_INTERVAL
+        return suggested
+
+    async def _handle_terminal_task(
+        self,
+        tool_name: str,
+        task_id: str,
+        snapshot: types.GetTaskResult,
+        parser: Callable[[types.CallToolResult], str | list[Content]],
+    ) -> str | list[Content]:
+        """Map a terminal task snapshot to either a parsed result or an exception."""
+        status = snapshot.status
+        if status == "completed":
+            payload = await self._fetch_task_result(task_id)
+            if payload.isError:
+                parsed = parser(payload)
+                text = (
+                    "\n".join(c.text for c in parsed if c.type == "text" and c.text)
+                    if isinstance(parsed, list)
+                    else str(parsed)
+                )
+                raise ToolExecutionException(text or str(parsed))
+            return parser(payload)
+
+        # Non-completed terminal statuses surface as ToolExecutionException so the
+        # function-calling loop sees a normal failure for tool_name.
+        message = snapshot.statusMessage or f"MCP task ended with status '{status}'."
+        if status == "input_required":
+            # Spec-non-terminal; treated as terminal here because the framework does
+            # not implement the interactive input flow.
+            message = snapshot.statusMessage or "MCP task requires additional input and cannot continue."
+        raise ToolExecutionException(f"Tool '{tool_name}' task {status}: {message}")
+
+    async def _fetch_task_result(self, task_id: str) -> types.CallToolResult:
+        """Send ``tasks/result`` and reinterpret the open-typed payload as a CallToolResult."""
+        from mcp import types
+        from mcp.shared.exceptions import McpError
+        from pydantic import ValidationError
+
+        request = types.ClientRequest(
+            types.GetTaskPayloadRequest(params=types.GetTaskPayloadRequestParams(taskId=task_id))
+        )
+        # Connection-loss retry only via the helper; no transient-code retry — server
+        # has already completed the task, so a slow payload fetch is anomalous.
+        try:
+            payload = await self._send_with_one_reconnect(
+                request, types.GetTaskPayloadResult, operation="tasks/result", task_id=task_id
+            )
+        except McpError as ex:
+            # Server reported completed; a hard fetch error is a plain failure (no cancel).
+            raise ToolExecutionException(ex.error.message, inner_exception=ex) from ex
+
+        # GetTaskPayloadResult carries the tool result via extra fields; reinterpret as CallToolResult.
+        payload_dict = payload.model_dump(by_alias=True, exclude_none=True)
+        payload_dict.pop("_meta", None)
+        try:
+            return types.CallToolResult.model_validate(payload_dict)
+        except ValidationError as ex:
+            # Server reported completed; malformed payload is a plain failure (no cancel needed).
+            raise ToolExecutionException(
+                f"MCP task '{task_id}' result payload could not be parsed as a CallToolResult.",
+                inner_exception=ex,
+            ) from ex
+
+    async def _send_with_one_reconnect(
+        self,
+        request: types.ClientRequest,
+        result_type: type[Any],
+        *,
+        operation: str,
+        task_id: str,
+    ) -> Any:
+        """Send ``request`` with one reconnect-and-retry on connection loss.
+
+        After a second loss (or reconnect failure), raise ``_MCPTaskAbandoned``.
+        Non-connection errors propagate unchanged.
+        """
+        from anyio import ClosedResourceError
+        from mcp.shared.exceptions import McpError
+
+        for attempt in range(_MCP_RECONNECT_ATTEMPTS):
+            try:
+                return await self.session.send_request(request, result_type)  # type: ignore[union-attr]
+            except (ClosedResourceError, McpError) as ex:
+                if not self._is_connection_lost(ex):
+                    raise
+                if attempt < _MCP_RECONNECT_ATTEMPTS - 1:
+                    logger.info(
+                        "MCP connection lost during %s; reconnecting (task_id=%s).", operation, task_id
+                    )
+                    try:
+                        await self.connect(reset=True)
+                    except Exception as reconn_ex:
+                        # Reconnect failure: task may still be running.
+                        raise _MCPTaskAbandoned(
+                            "Failed to reconnect to MCP server.", inner_exception=reconn_ex
+                        ) from reconn_ex
+                    continue
+                # Final attempt also lost the connection: task may still be running.
+                raise _MCPTaskAbandoned(
+                    f"MCP connection lost; task state unknown (task_id={task_id}).",
+                    inner_exception=ex,
+                ) from ex
+        raise AssertionError(f"unreachable: {operation} for {task_id}")  # pragma: no cover
+
+    @staticmethod
+    async def _await_with_deadline(coro: Coroutine[Any, Any, Any], timeout_s: float) -> Any:
+        """Await ``coro`` with a deadline; raise ``_MCPDeadlineExpired`` only on deadline.
+
+        Unlike ``asyncio.wait_for``, an ``asyncio.TimeoutError`` raised by ``coro``
+        itself propagates unchanged so callers can distinguish their own deadline
+        from a stray inner timeout.
+        """
+        inner = asyncio.ensure_future(coro)
+        try:
+            done, _pending = await asyncio.wait({inner}, timeout=timeout_s)
+        except BaseException:
+            # Outer caller cancelled (or another exception): cancel inner + drain.
+            inner.cancel()
+            with contextlib.suppress(BaseException):
+                await inner
+            raise
+        if inner in done:
+            return inner.result()
+        # Deadline fired before inner finished.
+        inner.cancel()
+        with contextlib.suppress(BaseException):
+            await inner
+        raise _MCPDeadlineExpired
+
+    def _spawn_best_effort_cancel(self, task_id: str) -> None:
+        """Fire-and-forget ``tasks/cancel`` so local cancellation propagates server-side."""
+        try:
+            loop = asyncio.get_running_loop()
+        except RuntimeError:
+            return
+        cancel_task = loop.create_task(self._try_cancel_task(task_id))
+        # Reuse pending-reload bookkeeping so close-on-owner waits/cancels these too.
+        self._pending_reload_tasks.add(cancel_task)
+        cancel_task.add_done_callback(self._pending_reload_tasks.discard)
+
+    async def _try_cancel_task(self, task_id: str) -> None:
+        """Send ``tasks/cancel``; bounded by ``_MCP_TASK_CANCEL_TIMEOUT``.
+
+        Failures log at warning so unattributed orphan tasks are debuggable.
+        """
+        from mcp import types
+
+        request = types.ClientRequest(
+            types.CancelTaskRequest(params=types.CancelTaskRequestParams(taskId=task_id))
+        )
+        try:
+            await asyncio.wait_for(
+                self.session.send_request(request, types.CancelTaskResult),  # type: ignore[union-attr]
+                timeout=_MCP_TASK_CANCEL_TIMEOUT.total_seconds(),
+            )
+        except asyncio.CancelledError:
+            raise
+        except asyncio.TimeoutError:
+            logger.warning(
+                "Best-effort tasks/cancel for '%s' timed out after %.1fs; "
+                "remote task may still be running.",
+                task_id,
+                _MCP_TASK_CANCEL_TIMEOUT.total_seconds(),
+            )
+        except Exception:
+            logger.warning(
+                "Best-effort tasks/cancel for '%s' failed; remote task may still be running.",
+                task_id,
+                exc_info=True,
+            )
+
+    @staticmethod
+    def _is_connection_lost(ex: BaseException) -> bool:
+        """Return True if *ex* indicates the MCP transport was torn down."""
+        from anyio import ClosedResourceError
+        from mcp.shared.exceptions import McpError
+
+        if isinstance(ex, ClosedResourceError):
+            return True
+        if isinstance(ex, McpError):
+            return "session terminated" in ex.error.message.lower()
+        return False
+
     async def get_prompt(self, prompt_name: str, **kwargs: Any) -> str:
         """Call a prompt with the given arguments.
 
@@ -1554,6 +2100,7 @@ class MCPStdioTool(MCPTool):
         encoding: str | None = None,
         client: SupportsChatGetResponse | None = None,
         additional_properties: dict[str, Any] | None = None,
+        task_options: MCPTaskOptions | None = None,
         **kwargs: Any,
     ) -> None:
         """Initialize the MCP stdio tool.
@@ -1598,6 +2145,8 @@ class MCPStdioTool(MCPTool):
             env: The environment variables to set for the command.
             encoding: The encoding to use for the command output.
             client: The chat client to use for sampling.
+            task_options: Options for tools that advertise
+                ``execution.taskSupport == "required"``. See :class:`MCPTaskOptions`.
             kwargs: Any extra arguments to pass to the stdio client.
         """
         super().__init__(
@@ -1614,6 +2163,7 @@ class MCPStdioTool(MCPTool):
             load_prompts=load_prompts,
             parse_prompt_results=parse_prompt_results,
             request_timeout=request_timeout,
+            task_options=task_options,
         )
         self.command = command
         self.args = args or []
@@ -1687,6 +2237,7 @@ class MCPStreamableHTTPTool(MCPTool):
         additional_properties: dict[str, Any] | None = None,
         http_client: AsyncClient | None = None,
         header_provider: Callable[[dict[str, Any]], dict[str, str]] | None = None,
+        task_options: MCPTaskOptions | None = None,
         **kwargs: Any,
     ) -> None:
         """Initialize the MCP streamable HTTP tool.
@@ -1739,6 +2290,8 @@ class MCPStreamableHTTPTool(MCPTool):
                 of HTTP headers to inject into every outbound request to the MCP server.
                 Use this to forward per-request context (e.g. authentication tokens set in
                 agent middleware) without creating a separate ``httpx.AsyncClient``.
+            task_options: Options for tools that advertise
+                ``execution.taskSupport == "required"``. See :class:`MCPTaskOptions`.
             kwargs: Additional keyword arguments (accepted for backward compatibility but not used).
         """
         super().__init__(
@@ -1755,6 +2308,7 @@ class MCPStreamableHTTPTool(MCPTool):
             load_prompts=load_prompts,
             parse_prompt_results=parse_prompt_results,
             request_timeout=request_timeout,
+            task_options=task_options,
         )
         self.url = url
         self.terminate_on_close = terminate_on_close
@@ -1862,6 +2416,7 @@ class MCPWebsocketTool(MCPTool):
         allowed_tools: Collection[str] | None = None,
         client: SupportsChatGetResponse | None = None,
         additional_properties: dict[str, Any] | None = None,
+        task_options: MCPTaskOptions | None = None,
         **kwargs: Any,
     ) -> None:
         """Initialize the MCP WebSocket tool.
@@ -1904,6 +2459,8 @@ class MCPWebsocketTool(MCPTool):
             allowed_tools: A list of tools that are allowed to use this tool.
             additional_properties: Additional properties.
             client: The chat client to use for sampling.
+            task_options: Options for tools that advertise
+                ``execution.taskSupport == "required"``. See :class:`MCPTaskOptions`.
             kwargs: Any extra arguments to pass to the WebSocket client.
         """
         super().__init__(
@@ -1920,6 +2477,7 @@ class MCPWebsocketTool(MCPTool):
             load_prompts=load_prompts,
             parse_prompt_results=parse_prompt_results,
             request_timeout=request_timeout,
+            task_options=task_options,
         )
         self.url = url
         self._client_kwargs = kwargs

python/packages/core/agent_framework/_middleware.py

@@ -11,6 +11,7 @@ from enum import Enum
 from typing import TYPE_CHECKING, Any, Generic, Literal, TypeAlias, cast, overload
 
 from ._clients import SupportsChatGetResponse
+from ._feature_stage import ExperimentalFeature, experimental
 from ._types import (
     AgentResponse,
     AgentResponseUpdate,
@@ -36,11 +37,10 @@ if TYPE_CHECKING:
     from pydantic import BaseModel
 
     from ._agents import SupportsAgentRun
-    from ._clients import SupportsChatGetResponse
     from ._compaction import CompactionStrategy, TokenizerProtocol
     from ._sessions import AgentSession
     from ._tools import FunctionTool, ToolTypes
-    from ._types import ChatOptions, ChatResponse, ChatResponseUpdate
+    from ._types import ChatOptions
 
     ResponseModelBoundT = TypeVar("ResponseModelBoundT", bound=BaseModel)
 
@@ -215,6 +215,12 @@ class FunctionInvocationContext:
         result: Function execution result. Can be observed after calling ``call_next()``
                 to see the actual execution result or can be set to override the execution result.
         kwargs: Additional runtime keyword arguments forwarded to the function invocation.
+        tools: The live, mutable list of tools available to the model for the current
+                agent run, or ``None`` when the function is invoked outside of a
+                function-calling loop (for example via ``FunctionTool.invoke`` directly).
+                Tools can add or remove tools during execution using :meth:`add_tools`
+                and :meth:`remove_tools` (progressive tool exposure). Mutations take
+                effect on the **next** model iteration, not the in-flight batch.
 
     Examples:
         .. code-block:: python
@@ -233,6 +239,18 @@ class FunctionInvocationContext:
 
                     # Continue execution
                     await call_next()
+
+        Progressive tool exposure from inside a tool:
+
+        .. code-block:: python
+
+            from agent_framework import FunctionInvocationContext, tool
+
+
+            @tool(approval_mode="never_require")
+            def load_math_tools(ctx: FunctionInvocationContext) -> str:
+                ctx.add_tools([factorial, fibonacci])
+                return "Math tools are now available."
     """
 
     def __init__(
@@ -243,6 +261,7 @@ class FunctionInvocationContext:
         metadata: Mapping[str, Any] | None = None,
         result: Any = None,
         kwargs: Mapping[str, Any] | None = None,
+        tools: list[ToolTypes] | None = None,
     ) -> None:
         """Initialize the FunctionInvocationContext.
 
@@ -253,6 +272,9 @@ class FunctionInvocationContext:
             metadata: Metadata dictionary for sharing data between function middleware.
             result: Function execution result.
             kwargs: Additional runtime keyword arguments forwarded to the function invocation.
+            tools: The live, mutable list of tools for the current agent run. When provided,
+                this is the same list object the model sees on the next iteration, so
+                appending or removing tools changes the model's available tools.
         """
         self.function = function
         self.arguments = arguments
@@ -260,6 +282,96 @@ class FunctionInvocationContext:
         self.metadata: dict[str, Any] = dict(metadata) if metadata is not None else {}
         self.result = result
         self.kwargs: dict[str, Any] = dict(kwargs) if kwargs is not None else {}
+        self.tools = tools
+
+    @experimental(feature_id=ExperimentalFeature.PROGRESSIVE_TOOLS)
+    def add_tools(
+        self,
+        tools: ToolTypes | Callable[..., Any] | Sequence[ToolTypes | Callable[..., Any]],
+    ) -> None:
+        """Add one or more tools to the current agent run (progressive tool exposure).
+
+        Callable inputs are converted to :class:`FunctionTool`, and tool collections are
+        flattened, using the same normalization as the rest of the framework. Added tools
+        become available to the model on the **next** iteration of the function-calling
+        loop; they do not affect tool calls already requested in the in-flight batch.
+
+        Adding a tool whose name already exists is a no-op when it is the same object, and
+        raises ``ValueError`` when it is a different object with a duplicate name.
+
+        Args:
+            tools: A single tool/callable or a sequence of tools/callables to add.
+
+        Raises:
+            RuntimeError: If the context has no live tools list (for example when the
+                function is invoked outside of a function-calling loop).
+            ValueError: If a different tool with a duplicate name is added.
+        """
+        from ._tools import _append_unique_tools, normalize_tools  # type: ignore[reportPrivateUsage]
+
+        if self.tools is None:
+            raise RuntimeError(
+                "Cannot add tools: this FunctionInvocationContext is not bound to a live "
+                "agent run. add_tools is only available for functions invoked within an "
+                "agent's function-calling loop."
+            )
+        # Validate the whole batch against a throwaway copy first, so a duplicate-name
+        # clash partway through the batch raises before the live tool list is mutated
+        # (all-or-nothing semantics).
+        merged = _append_unique_tools(list(self.tools), normalize_tools(tools))
+        self.tools[:] = merged
+
+    @experimental(feature_id=ExperimentalFeature.PROGRESSIVE_TOOLS)
+    def remove_tools(
+        self,
+        tools: ToolTypes | Callable[..., Any] | Sequence[ToolTypes | Callable[..., Any]] | str | Sequence[str],
+    ) -> None:
+        """Remove one or more tools from the current agent run (progressive tool exposure).
+
+        Tools may be specified by name, by tool object, or by the original callable. Names
+        that are not currently present are ignored. Removals take effect on the **next**
+        iteration of the function-calling loop; tool calls already requested in the
+        in-flight batch still execute.
+
+        Args:
+            tools: A tool name, tool/callable, or a sequence of any of these to remove.
+
+        Raises:
+            RuntimeError: If the context has no live tools list (for example when the
+                function is invoked outside of a function-calling loop).
+        """
+        from ._tools import _get_tool_name, normalize_tools  # type: ignore[reportPrivateUsage]
+
+        if self.tools is None:
+            raise RuntimeError(
+                "Cannot remove tools: this FunctionInvocationContext is not bound to a live "
+                "agent run. remove_tools is only available for functions invoked within an "
+                "agent's function-calling loop."
+            )
+
+        names_to_remove: set[str] = set()
+        raw_items: list[Any]
+        if isinstance(tools, str):
+            raw_items = [tools]
+        elif isinstance(tools, Sequence) and not isinstance(tools, (bytes, bytearray)):
+            raw_items = list(cast("Sequence[Any]", tools))
+        else:
+            raw_items = [tools]
+        for item in raw_items:
+            if isinstance(item, str):
+                names_to_remove.add(item)
+                continue
+            for normalized in normalize_tools(item):
+                if name := _get_tool_name(normalized):  # type: ignore[reportPrivateUsage]
+                    names_to_remove.add(name)
+
+        if not names_to_remove:
+            return
+        self.tools[:] = [
+            tool
+            for tool in self.tools
+            if _get_tool_name(tool) not in names_to_remove  # type: ignore[reportPrivateUsage]
+        ]
 
 
 class ChatContext:

python/packages/core/agent_framework/_serialization.py

@@ -7,6 +7,8 @@ import json
 import logging
 import re
 from collections.abc import Mapping, MutableMapping
+from dataclasses import asdict, is_dataclass
+from datetime import date, datetime
 from typing import Any, ClassVar, Protocol, TypeVar, runtime_checkable
 
 logger = logging.getLogger("agent_framework")
@@ -614,3 +616,46 @@ class SerializationMixin:
         # Fallback and default
         # Convert class name to snake_case
         return _CAMEL_TO_SNAKE_PATTERN.sub("_", cls.__name__).lower()
+
+
+def make_json_safe(obj: Any) -> Any:
+    """Recursively convert an object to a JSON-serializable form.
+
+    Handles dataclasses, Pydantic models, objects with ``to_dict``/``dict``/``__dict__``,
+    datetimes, lists, dicts, and primitives.  Falls back to ``str()`` for any remaining
+    non-serializable value so that ``json.dumps`` never raises a ``TypeError``.
+
+    Args:
+        obj: Object to make JSON safe.
+
+    Returns:
+        A JSON-serializable version of the object.
+    """
+    if obj is None or isinstance(obj, (str, int, float, bool)):
+        return obj
+    if isinstance(obj, (datetime, date)):
+        return obj.isoformat()
+    if is_dataclass(obj) and not isinstance(obj, type):
+        return make_json_safe(asdict(obj))  # type: ignore[arg-type]
+    if callable(getattr(obj, "model_dump", None)):
+        try:
+            return make_json_safe(obj.model_dump())  # type: ignore[no-any-return]
+        except TypeError:
+            pass
+    if callable(getattr(obj, "to_dict", None)):
+        try:
+            return make_json_safe(obj.to_dict())  # type: ignore[no-any-return]
+        except TypeError:
+            pass
+    if callable(getattr(obj, "dict", None)):
+        try:
+            return make_json_safe(obj.dict())  # type: ignore[no-any-return]
+        except TypeError:
+            pass
+    if isinstance(obj, dict):
+        return {str(key): make_json_safe(value) for key, value in obj.items()}  # type: ignore[misc]
+    if isinstance(obj, (list, tuple)):
+        return [make_json_safe(item) for item in obj]  # type: ignore[misc]
+    if hasattr(obj, "__dict__"):
+        return {key: make_json_safe(value) for key, value in vars(obj).items()}  # type: ignore[misc]
+    return str(obj)

python/packages/core/agent_framework/_skills.py

@@ -44,6 +44,7 @@ Only use skills from trusted sources.
 from __future__ import annotations
 
 import asyncio
+import base64
 import inspect
 import json
 import logging
@@ -60,6 +61,10 @@ from ._sessions import ContextProvider
 from ._tools import FunctionTool
 
 if TYPE_CHECKING:
+    from mcp.client.session import ClientSession
+    from mcp.types import ReadResourceResult
+    from pydantic import AnyUrl
+
     from ._agents import SupportsAgentRun
     from ._sessions import AgentSession, SessionContext
 
@@ -2134,9 +2139,7 @@ class SkillsProvider(ContextProvider):
             ),
             FunctionTool(
                 name="read_skill_resource",
-                description=(
-                    "Reads a resource associated with a skill, such as references, assets, or dynamic data."
-                ),
+                description=("Reads a resource associated with a skill, such as references, assets, or dynamic data."),
                 func=_read_resource,
                 input_model={
                     "type": "object",
@@ -2173,8 +2176,7 @@ class SkillsProvider(ContextProvider):
                                     "type": "object",
                                     "additionalProperties": True,
                                     "description": (
-                                        "Named arguments as key-value pairs "
-                                        '(e.g. {"length": 24, "uppercase": true}).'
+                                        'Named arguments as key-value pairs (e.g. {"length": 24, "uppercase": true}).'
                                     ),
                                 },
                                 {
@@ -3288,4 +3290,443 @@ class AggregatingSkillsSource(SkillsSource):
         return result
 
 
+# region MCP Skills
+
+
+def _mcp_any_url(uri: str) -> AnyUrl:
+    """Convert a string URI to a :class:`pydantic.AnyUrl` for MCP client calls."""
+    from pydantic import AnyUrl as _AnyUrl
+
+    return _AnyUrl(uri)
+
+
+def _is_mcp_resource_not_found(ex: Exception) -> bool:
+    """Return ``True`` when *ex* is an :class:`McpError` indicating a missing resource.
+
+    Two codes are treated as "not found":
+
+    * ``-32002`` — the MCP-spec "Resource not found" code returned by a
+      compliant server when the URI does not exist. Not exported as a
+      constant from ``mcp.types`` but defined by the resources subprotocol.
+    * ``METHOD_NOT_FOUND`` (``-32601``) — the server does not implement
+      ``resources/read`` at all, which for the skills source is functionally
+      equivalent to "no skills available."
+
+    All other codes — ``INVALID_PARAMS``, ``INTERNAL_ERROR``, ``PARSE_ERROR``,
+    ``CONNECTION_CLOSED``, auth rejections, and generic handler errors
+    (code ``0``) — are treated as real failures so that a misconfigured
+    token or crashing server is not silently mistaken for "the server has no
+    skills."
+    """
+    from mcp.shared.exceptions import McpError as _McpError
+
+    if not isinstance(ex, _McpError):
+        return False
+    from mcp.types import METHOD_NOT_FOUND as _METHOD_NOT_FOUND
+
+    return ex.error.code in {-32002, _METHOD_NOT_FOUND}
+
+
+def _mcp_join_text(result: ReadResourceResult) -> str:
+    """Join all :class:`TextResourceContents` items in a result into a single string."""
+    from mcp.types import TextResourceContents as _TextResourceContents
+
+    return "\n".join(c.text for c in result.contents if isinstance(c, _TextResourceContents))
+
+
+class _McpSkillIndexEntry:  # noqa: B903
+    """A single entry in the ``skill://index.json`` discovery document.
+
+    All fields are optional to support lenient deserialization; callers
+    validate required fields before use.
+    """
+
+    def __init__(
+        self,
+        *,
+        name: str | None = None,
+        type: str | None = None,
+        description: str | None = None,
+        url: str | None = None,
+        digest: str | None = None,
+    ) -> None:
+        self.name = name
+        self.type = type
+        self.description = description
+        self.url = url
+        self.digest = digest
+
+
+class _McpSkillIndex:
+    """DTO for the ``skill://index.json`` discovery document.
+
+    Represents the Agent Skills Discovery v0.2.0 schema as bound to MCP
+    by SEP-2640.
+    """
+
+    def __init__(
+        self,
+        *,
+        schema: str | None = None,
+        skills: list[_McpSkillIndexEntry] | None = None,
+    ) -> None:
+        self.schema = schema
+        self.skills: list[_McpSkillIndexEntry] = skills if skills is not None else []
+
+
+def _parse_mcp_skill_index(text: str) -> _McpSkillIndex:
+    """Parse a JSON string into a :class:`_McpSkillIndex`.
+
+    Args:
+        text: Raw JSON text from ``skill://index.json``.
+
+    Returns:
+        A populated :class:`_McpSkillIndex` instance.
+
+    Raises:
+        json.JSONDecodeError: If the text is not valid JSON.
+        ValueError: If the top-level value is not a JSON object.
+    """
+    raw: dict[str, Any] = json.loads(text)
+
+    if not isinstance(raw, dict):
+        raise ValueError("skill://index.json must be a JSON object")
+
+    entries: list[_McpSkillIndexEntry] = []
+
+    raw_skills: list[Any] = raw.get("skills") or []
+
+    for item in raw_skills:
+        if isinstance(item, dict):
+            d = cast(dict[str, Any], item)
+
+            entries.append(
+                _McpSkillIndexEntry(
+                    name=d.get("name"),
+                    type=d.get("type"),
+                    description=d.get("description"),
+                    url=d.get("url"),
+                    digest=d.get("digest"),
+                )
+            )
+
+    return _McpSkillIndex(schema=raw.get("$schema"), skills=entries)
+
+
+@experimental(feature_id=ExperimentalFeature.MCP_SKILLS)
+class MCPSkillResource(SkillResource):
+    """A :class:`SkillResource` backed by content fetched from an MCP server.
+
+    The :class:`~mcp.types.ReadResourceResult` is fetched eagerly by
+    :meth:`MCPSkill.get_resource` at construction time; :meth:`read`
+    extracts text or binary content from the result.
+    """
+
+    def __init__(self, *, name: str, result: ReadResourceResult) -> None:
+        """Initialize an MCPSkillResource.
+
+        Args:
+            name: The resource name (e.g. a relative path or identifier).
+            result: The result returned by the MCP server's ``resources/read`` request.
+        """
+        super().__init__(name=name)
+        self._result = result
+
+    async def read(self, **kwargs: Any) -> Any:
+        """Read the resource content.
+
+        Returns:
+            A ``bytes`` object when the resource contains binary content,
+            a ``str`` when it contains text, or ``None`` when the server
+            returned no content blocks.
+        """
+        from mcp.types import BlobResourceContents, TextResourceContents
+
+        for content in self._result.contents:
+            if isinstance(content, BlobResourceContents):
+                blob = content.blob
+                # Strip data-URI prefix if present (some MCP servers send
+                # full data URIs instead of raw base64).
+                if blob.startswith("data:"):
+                    blob = blob.split(",", 1)[-1]
+                return base64.b64decode(blob)
+
+        text = "\n".join(c.text for c in self._result.contents if isinstance(c, TextResourceContents))
+        return text if text else None
+
+
+@experimental(feature_id=ExperimentalFeature.MCP_SKILLS)
+class MCPSkill(Skill):
+    """A :class:`Skill` discovered from an MCP server exposing the Agent Skills convention.
+
+    The skill is constructed from ``skill://index.json`` discovery metadata;
+    :meth:`get_content` fetches the full ``SKILL.md`` content from the MCP
+    server on demand via ``resources/read``.
+
+    Per SEP-2640, resources referenced inside SKILL.md are fetched on demand
+     via the originating MCP server: :meth:`get_resource` resolves a relative
+    resource name against the skill's root URI, issues a ``resources/read``
+     request, and returns an :class:`MCPSkillResource` with pre-fetched content.
+    """
+
+    _SKILL_MD_SUFFIX: Final[str] = "SKILL.md"
+
+    def __init__(
+        self,
+        frontmatter: SkillFrontmatter,
+        skill_md_uri: str,
+        client: ClientSession,
+    ) -> None:
+        """Initialize an MCPSkill.
+
+        Args:
+            frontmatter: The parsed frontmatter metadata for this skill.
+            skill_md_uri: The full MCP resource URI of the ``SKILL.md`` resource
+                (e.g. ``skill://unit-converter/SKILL.md``). The skill's root URI
+                is derived by stripping the trailing ``SKILL.md`` segment.
+            client: The MCP client session used to fetch resources on demand.
+        """
+        self._frontmatter = frontmatter
+        self._skill_md_uri = skill_md_uri
+        self._skill_root_uri = self._compute_skill_root_uri(skill_md_uri)
+        self._client = client
+        self._content: str | None = None
+
+    @property
+    def frontmatter(self) -> SkillFrontmatter:
+        """The L1 discovery metadata for this skill."""
+        return self._frontmatter
+
+    async def get_content(self) -> str:
+        """Get the full SKILL.md content from the MCP server.
+
+        Fetches the content via ``resources/read`` on the first call and
+        caches the result for subsequent calls.
+
+        Returns:
+            The SKILL.md content string.
+
+        Raises:
+            ValueError: If the MCP server returned no text content for the
+                SKILL.md resource.
+        """
+        if self._content is not None:
+            return self._content
+
+        result = await self._client.read_resource(_mcp_any_url(self._skill_md_uri))
+        text = _mcp_join_text(result)
+        if not text:
+            raise ValueError(
+                f"The MCP server returned no text content for SKILL.md resource '{self._skill_md_uri}'."
+            )
+        self._content = text
+        return text
+
+    async def get_resource(self, name: str) -> SkillResource | None:
+        """Get a sibling resource by name from the MCP server.
+
+        Resolves *name* as a relative path against the skill's root URI,
+        issues a ``resources/read`` request to the MCP server, and returns
+        an :class:`MCPSkillResource` with the pre-fetched content.
+
+        Args:
+            name: The resource name (e.g. ``references/checklist.md``).
+
+        Returns:
+            An :class:`MCPSkillResource`, or ``None`` when the name is empty
+            or the resource does not exist on the server.
+        """
+        if not name or not name.strip():
+            return None
+
+        normalized = self._validate_resource_name(name)
+        if normalized is None:
+            return None
+
+        uri = self._skill_root_uri + normalized
+        try:
+            result = await self._client.read_resource(_mcp_any_url(uri))
+        except Exception as ex:
+            if _is_mcp_resource_not_found(ex):
+                logger.debug("MCP resource '%s' not available: %s", uri, ex)
+                return None
+            raise
+
+        return MCPSkillResource(name=name, result=result)
+
+    @staticmethod
+    def _validate_resource_name(name: str) -> str | None:
+        """Validate a resource name and return the normalized form.
+
+        Defense in depth: refuses names that could escape the skill root
+        (absolute paths, embedded URI schemes, parent-traversal segments).
+        The MCP server is the authority on URI resolution, but rejecting
+        obviously unsafe shapes client-side avoids leaking escape attempts
+        upstream.
+
+        Args:
+            name: The raw resource name to validate.
+
+        Returns:
+            The normalized name with backslashes replaced by forward slashes,
+            or ``None`` if the name is unsafe.
+        """
+        normalized = name.replace("\\", "/")
+        if (
+            normalized.startswith("/")
+            or "://" in normalized
+            or any(seg == ".." for seg in normalized.split("/"))
+        ):
+            logger.debug("Rejecting resource name with unsafe path components: %r", name)
+            return None
+        return normalized
+
+    @staticmethod
+    def _compute_skill_root_uri(skill_md_uri: str) -> str:
+        """Strip the trailing ``SKILL.md`` from the URI to produce the skill root.
+
+        If the URI doesn't end with ``SKILL.md``, ensures it ends with a
+        trailing slash.
+        """
+        if skill_md_uri.endswith(MCPSkill._SKILL_MD_SUFFIX):
+            return skill_md_uri[: -len(MCPSkill._SKILL_MD_SUFFIX)]
+        if skill_md_uri.endswith("/"):
+            return skill_md_uri
+        return skill_md_uri + "/"
+
+
+@experimental(feature_id=ExperimentalFeature.MCP_SKILLS)
+class MCPSkillsSource(SkillsSource):
+    """A :class:`SkillsSource` that discovers Agent Skills served over MCP.
+
+    Discovery follows the SEP-2640 recommended approach: the source reads
+    the well-known ``skill://index.json`` resource and constructs one
+    :class:`MCPSkill` per ``skill-md`` entry directly from the entry's
+    ``name``, ``description``, and ``url`` fields.
+
+    The referenced ``SKILL.md`` resource is **not** read during discovery;
+    the host fetches its body on demand via ``resources/read`` when the
+    skill content is needed.
+
+    Only index entries of type ``skill-md`` are supported; entries of any
+    other type are silently skipped.
+
+    If ``skill://index.json`` is absent, unreadable, empty, or fails to
+    parse, this source returns an empty list.
+
+    Examples:
+        .. code-block:: python
+
+            from mcp.client.session import ClientSession
+
+            source = MCPSkillsSource(client=session)
+            skills = await source.get_skills()
+    """
+
+    _INDEX_URI: Final[str] = "skill://index.json"
+    _SKILL_MD_TYPE: Final[str] = "skill-md"
+
+    def __init__(self, client: ClientSession) -> None:
+        """Initialize an MCPSkillsSource.
+
+        Args:
+            client: An MCP client session connected to a server that
+                exposes Agent Skills resources.
+        """
+        self._client = client
+
+    async def get_skills(self) -> list[Skill]:
+        """Discover and return skills from the MCP server.
+
+        Reads ``skill://index.json``, parses it, and creates an
+        :class:`MCPSkill` for each valid ``skill-md`` entry.
+
+        Returns:
+            A list of discovered :class:`MCPSkill` instances.
+        """
+        index = await self._try_read_index()
+        if index is None:
+            return []
+
+        skills: list[Skill] = []
+        for entry in index.skills:
+            result = self._try_create_skill(entry)
+            if result is not None:
+                skills.append(result)
+                logger.info("Loaded MCP skill: %s", result.frontmatter.name)
+            else:
+                logger.debug(
+                    "Skipping skill index entry '%s'",
+                    entry.name or "(unnamed)",
+                )
+
+        logger.info("Successfully loaded %d skills from MCP server", len(skills))
+        return skills
+
+    async def _try_read_index(self) -> _McpSkillIndex | None:
+        """Attempt to read and parse ``skill://index.json`` from the MCP server.
+
+        Returns:
+            A parsed :class:`_McpSkillIndex`, or ``None`` if the index is
+            absent, empty, or malformed.
+        """
+        try:
+            result = await self._client.read_resource(_mcp_any_url(self._INDEX_URI))
+        except Exception as ex:
+            if _is_mcp_resource_not_found(ex):
+                logger.debug("No skill://index.json resource available on MCP server: %s", ex)
+                return None
+            logger.warning("Failed to read skill://index.json from MCP server.", exc_info=True)
+            raise
+
+        index_text = _mcp_join_text(result)
+        if not index_text:
+            logger.debug("skill://index.json on MCP server returned empty/non-text contents")
+            return None
+
+        try:
+            return _parse_mcp_skill_index(index_text)
+        except (json.JSONDecodeError, ValueError):
+            logger.warning("Failed to parse skill://index.json JSON document.", exc_info=True)
+            return None
+
+    def _try_create_skill(self, entry: _McpSkillIndexEntry) -> MCPSkill | None:
+        """Attempt to create an :class:`MCPSkill` from an index entry.
+
+        Args:
+            entry: A single entry from the skill index.
+
+        Returns:
+            An :class:`MCPSkill` if the entry is valid, or ``None`` if the
+            entry should be skipped.
+        """
+        if entry.type != self._SKILL_MD_TYPE:
+            logger.debug(
+                "Skipping entry '%s': unsupported type '%s'",
+                entry.name or "(unnamed)",
+                entry.type or "(none)",
+            )
+            return None
+
+        if not entry.name or not entry.name.strip():
+            logger.debug("Skipping entry: missing required 'name' field")
+            return None
+
+        if not entry.description or not entry.description.strip():
+            logger.debug("Skipping entry '%s': missing required 'description' field", entry.name)
+            return None
+
+        if not entry.url or not entry.url.strip():
+            logger.debug("Skipping entry '%s': missing required 'url' field", entry.name)
+            return None
+
+        try:
+            fm = SkillFrontmatter(name=entry.name, description=entry.description)
+        except ValueError as ex:
+            logger.debug("Skipping entry '%s': invalid metadata: %s", entry.name, ex)
+            return None
+
+        return MCPSkill(frontmatter=fm, skill_md_uri=entry.url, client=self._client)
+
+
 # endregion

python/packages/core/agent_framework/_tools.py

@@ -1418,6 +1418,7 @@ async def _auto_invoke_function(
     sequence_index: int | None = None,
     request_index: int | None = None,
     middleware_pipeline: FunctionMiddlewarePipeline | None = None,
+    live_tools: list[ToolTypes] | None = None,
 ) -> Content:
     """Invoke a function call requested by the agent, applying middleware that is defined.
 
@@ -1432,6 +1433,8 @@ async def _auto_invoke_function(
         sequence_index: The index of the function call in the sequence.
         request_index: The index of the request iteration.
         middleware_pipeline: Optional middleware pipeline to apply during execution.
+        live_tools: The live, mutable tools list for the current agent run, exposed on
+            the FunctionInvocationContext so tools can add/remove tools at runtime.
 
     Returns:
         The function result content.
@@ -1523,6 +1526,7 @@ async def _auto_invoke_function(
                     arguments=args,
                     session=invocation_session,
                     kwargs=runtime_kwargs.copy(),
+                    tools=live_tools,
                 )
             function_result = await tool.invoke(
                 arguments=args,
@@ -1537,6 +1541,10 @@ async def _auto_invoke_function(
         except UserInputRequiredException:
             raise
         except Exception as exc:
+            logger.warning(
+                f"Function '{tool.name}' raised an exception; returning an error result to the "
+                f"model. Set include_detailed_errors=True for the full detail. Exception: {exc!r}"
+            )
             message = "Error: Function failed."
             if config.get("include_detailed_errors", False):
                 message = f"{message} Exception: {exc}"
@@ -1552,6 +1560,7 @@ async def _auto_invoke_function(
         arguments=args,
         session=invocation_session,
         kwargs=runtime_kwargs.copy(),
+        tools=live_tools,
     )
 
     call_id = function_call_content.call_id
@@ -1608,6 +1617,10 @@ async def _auto_invoke_function(
     except UserInputRequiredException:
         raise
     except Exception as exc:
+        logger.warning(
+            f"Function '{tool.name}' raised an exception; returning an error result to the "
+            f"model. Set include_detailed_errors=True for the full detail. Exception: {exc!r}"
+        )
         message = "Error: Function failed."
         if config.get("include_detailed_errors", False):
             message = f"{message} Exception: {exc}"
@@ -1659,6 +1672,9 @@ async def _try_execute_function_calls(
     from ._types import Content
 
     tool_map = _get_tool_map(tools)
+    # The live tools list (when tools is the run-local list) is exposed on the
+    # FunctionInvocationContext so tools can add/remove tools during the run.
+    live_tools: list[ToolTypes] | None = cast("list[ToolTypes]", tools) if isinstance(tools, list) else None
     approval_tools = [tool_name for tool_name, tool in tool_map.items() if tool.approval_mode == "always_require"]
     logger.debug(
         "_try_execute_function_calls: tool_map keys=%s, approval_tools=%s",
@@ -1733,6 +1749,7 @@ async def _try_execute_function_calls(
                 request_index=attempt_idx,
                 middleware_pipeline=middleware_pipeline,
                 config=config,
+                live_tools=live_tools,
             )
             return (result, False)
         except MiddlewareTermination as exc:
@@ -2371,6 +2388,13 @@ class FunctionInvocationLayer(Generic[OptionsCoT]):
                 function_invocation_kwargs=function_invocation_kwargs,
                 client_kwargs=filtered_kwargs,
             )
+        # Establish a single, run-local mutable tools list so that tools can add or remove
+        # tools during the run (progressive tool exposure). A fresh list is created via
+        # normalize_tools so the caller's original tools container is never mutated, while
+        # the same list object is shared with the model (options["tools"]) and the tool map
+        # rebuilt on every loop iteration.
+        if mutable_options.get("tools"):
+            mutable_options["tools"] = normalize_tools(mutable_options["tools"])
         if not stream:
 
             async def _get_response() -> ChatResponse[Any]:

python/packages/core/agent_framework/_workflows/_agent.py

@@ -12,6 +12,7 @@ from datetime import datetime, timezone
 from typing import TYPE_CHECKING, Any, ClassVar, Literal, cast, overload
 
 from .._agents import BaseAgent
+from .._serialization import make_json_safe
 from .._sessions import (
     AgentSession,
     ContextProvider,
@@ -61,7 +62,7 @@ class WorkflowAgent(BaseAgent):
         data: Any
 
         def to_dict(self) -> dict[str, Any]:
-            return {"request_id": self.request_id, "data": self.data}
+            return {"request_id": self.request_id, "data": make_json_safe(self.data)}
 
         def to_json(self) -> str:
             return json.dumps(self.to_dict())

python/packages/core/agent_framework/_workflows/_functional.py

@@ -47,6 +47,7 @@ from copy import deepcopy
 from typing import Any, Generic, Literal, TypeVar, overload
 
 from .._feature_stage import ExperimentalFeature, experimental
+from .._serialization import make_json_safe
 from .._types import AgentResponse, AgentResponseUpdate, ResponseStream
 from ..observability import OtelAttr, capture_exception, create_workflow_span
 from ._checkpoint import CheckpointStorage, WorkflowCheckpoint
@@ -1515,7 +1516,7 @@ class FunctionalWorkflowAgent:
         function_call = Content.from_function_call(
             call_id=request_id,
             name=self.REQUEST_INFO_FUNCTION_NAME,
-            arguments={"request_id": request_id, "data": event.data},
+            arguments={"request_id": request_id, "data": make_json_safe(event.data)},
         )
         return Content.from_function_approval_request(
             id=request_id,

python/packages/core/agent_framework/foundry/__init__.py

@@ -34,6 +34,7 @@ _IMPORTS: dict[str, tuple[str, str]] = {
     "FoundryLocalChatOptions": ("agent_framework_foundry_local", "agent-framework-foundry-local"),
     "FoundryLocalClient": ("agent_framework_foundry_local", "agent-framework-foundry-local"),
     "FoundryLocalSettings": ("agent_framework_foundry_local", "agent-framework-foundry-local"),
+    "GeneratedEvaluatorRef": ("agent_framework_foundry", "agent-framework-foundry"),
     "RawAnthropicFoundryClient": ("agent_framework_anthropic", "agent-framework-anthropic"),
     "RawFoundryAgent": ("agent_framework_foundry", "agent-framework-foundry"),
     "RawFoundryAgentChatClient": ("agent_framework_foundry", "agent-framework-foundry"),

python/packages/core/agent_framework/foundry/__init__.pyi

@@ -20,6 +20,7 @@ from agent_framework_foundry import (
     FoundryEmbeddingSettings,
     FoundryEvals,
     FoundryMemoryProvider,
+    GeneratedEvaluatorRef,
     RawFoundryAgent,
     RawFoundryAgentChatClient,
     RawFoundryChatClient,
@@ -52,6 +53,7 @@ __all__ = [
     "FoundryLocalClient",
     "FoundryLocalSettings",
     "FoundryMemoryProvider",
+    "GeneratedEvaluatorRef",
     "RawAnthropicFoundryClient",
     "RawFoundryAgent",
     "RawFoundryAgentChatClient",

python/packages/core/agent_framework/observability.py

@@ -498,14 +498,34 @@ def _get_exporters_from_env(
     # Get base endpoint
     base_endpoint = os.getenv("OTEL_EXPORTER_OTLP_ENDPOINT")
 
-    # Get signal-specific endpoints (these override base endpoint)
-    traces_endpoint = os.getenv("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT") or base_endpoint
-    metrics_endpoint = os.getenv("OTEL_EXPORTER_OTLP_METRICS_ENDPOINT") or base_endpoint
-    logs_endpoint = os.getenv("OTEL_EXPORTER_OTLP_LOGS_ENDPOINT") or base_endpoint
+    # Get signal-specific endpoints (these override base endpoint and are used verbatim)
+    traces_endpoint_specific = os.getenv("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT")
+    metrics_endpoint_specific = os.getenv("OTEL_EXPORTER_OTLP_METRICS_ENDPOINT")
+    logs_endpoint_specific = os.getenv("OTEL_EXPORTER_OTLP_LOGS_ENDPOINT")
 
     # Get protocol (default is grpc)
     protocol = os.getenv("OTEL_EXPORTER_OTLP_PROTOCOL", "grpc").lower()
 
+    # Per the OTel spec, OTEL_EXPORTER_OTLP_ENDPOINT is a *base* URL for HTTP — the SDK
+    # auto-appends /v1/{traces,metrics,logs} when it reads the env var directly. The
+    # signal-specific endpoint env vars are *full* URLs used verbatim. Because we read
+    # the env vars here and forward them as the ``endpoint=`` constructor argument
+    # (which the SDK always treats as a full URL), we must replicate the auto-append
+    # ourselves for HTTP when falling back to the base endpoint. For gRPC, the base
+    # endpoint is used as-is.
+    traces_endpoint: str | None
+    metrics_endpoint: str | None
+    logs_endpoint: str | None
+    if protocol in ("http/protobuf", "http") and base_endpoint:
+        base_for_http = base_endpoint.rstrip("/")
+        traces_endpoint = traces_endpoint_specific or f"{base_for_http}/v1/traces"
+        metrics_endpoint = metrics_endpoint_specific or f"{base_for_http}/v1/metrics"
+        logs_endpoint = logs_endpoint_specific or f"{base_for_http}/v1/logs"
+    else:
+        traces_endpoint = traces_endpoint_specific or base_endpoint
+        metrics_endpoint = metrics_endpoint_specific or base_endpoint
+        logs_endpoint = logs_endpoint_specific or base_endpoint
+
     # Get base headers
     base_headers_str = os.getenv("OTEL_EXPORTER_OTLP_HEADERS", "")
     base_headers = _parse_headers(base_headers_str)

python/packages/core/tests/core/test_function_invocation_logic.py

@@ -3975,3 +3975,425 @@ async def test_user_input_request_empty_contents_returns_fallback(chat_client_ba
     ]
     assert len(function_results) >= 1
     assert any("user input" in (fr.result or "").lower() for fr in function_results)
+
+
+# region Progressive tool exposure (FunctionInvocationContext.add_tools / remove_tools)
+
+
+def _pte_function_call_response(call_id: str, name: str, arguments: str = "{}") -> ChatResponse:
+    return ChatResponse(
+        messages=Message(
+            role="assistant",
+            contents=[Content.from_function_call(call_id=call_id, name=name, arguments=arguments)],
+        )
+    )
+
+
+def _pte_text_response(text: str = "done") -> ChatResponse:
+    return ChatResponse(messages=Message(role="assistant", contents=[text]))
+
+
+@tool(name="factorial", approval_mode="never_require")
+def _pte_factorial(n: int) -> int:
+    """Compute the factorial of n."""
+    result = 1
+    for value in range(2, n + 1):
+        result *= value
+    return result
+
+
+async def test_context_exposes_live_tools(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    seen_names: list[str] = []
+
+    @tool(name="inspect_tools", approval_mode="never_require")
+    def inspect_tools(ctx: FunctionInvocationContext) -> str:
+        assert ctx.tools is not None
+        seen_names.extend(t.name for t in ctx.tools if isinstance(t, FunctionTool))
+        return "inspected"
+
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "inspect_tools"),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [inspect_tools]},
+    )
+    assert "inspect_tools" in seen_names
+
+
+async def test_add_tools_available_next_iteration(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    @tool(name="factorial", approval_mode="never_require")
+    def factorial(n: int) -> int:
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(factorial)
+        return "math tools loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    response = await chat_client_base.get_response(
+        [Message(role="user", contents=["compute 5!"])],
+        options={"tool_choice": "auto", "tools": [load_math]},
+    )
+    assert exec_counter == 1
+    assert response.messages[-1].text == "done"
+
+
+async def test_add_tools_model_sees_added_tools_in_options(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    recorded: list[list[str]] = []
+    client_cls = type(chat_client_base)
+    original = client_cls._get_non_streaming_response
+
+    async def recording(self: Any, *, messages: Any, options: dict[str, Any], **kwargs: Any) -> ChatResponse:
+        tools = options.get("tools") or []
+        recorded.append([t.name for t in tools if isinstance(t, FunctionTool)])
+        return await original(self, messages=messages, options=options, **kwargs)
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(_pte_factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    monkey = pytest.MonkeyPatch()
+    monkey.setattr(client_cls, "_get_non_streaming_response", recording)
+    try:
+        await chat_client_base.get_response(
+            [Message(role="user", contents=["compute 5!"])],
+            options={"tool_choice": "auto", "tools": [load_math]},
+        )
+    finally:
+        monkey.undo()
+
+    assert recorded[0] == ["load_math"]
+    assert "factorial" in recorded[1]
+
+
+async def test_remove_tools_next_iteration(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    recorded: list[list[str]] = []
+    client_cls = type(chat_client_base)
+    original = client_cls._get_non_streaming_response
+
+    async def recording(self: Any, *, messages: Any, options: dict[str, Any], **kwargs: Any) -> ChatResponse:
+        tools = options.get("tools") or []
+        recorded.append([t.name for t in tools if isinstance(t, FunctionTool)])
+        return await original(self, messages=messages, options=options, **kwargs)
+
+    @tool(name="get_weather", approval_mode="never_require")
+    def get_weather(location: str) -> str:
+        return "sunny"
+
+    @tool(name="drop_weather", approval_mode="never_require")
+    def drop_weather(ctx: FunctionInvocationContext) -> str:
+        ctx.remove_tools("get_weather")
+        return "removed"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "drop_weather"),
+        _pte_text_response(),
+    ]
+    monkey = pytest.MonkeyPatch()
+    monkey.setattr(client_cls, "_get_non_streaming_response", recording)
+    try:
+        await chat_client_base.get_response(
+            [Message(role="user", contents=["hi"])],
+            options={"tool_choice": "auto", "tools": [get_weather, drop_weather]},
+        )
+    finally:
+        monkey.undo()
+
+    assert set(recorded[0]) == {"get_weather", "drop_weather"}
+    assert "get_weather" not in recorded[1]
+
+
+async def test_add_tools_does_not_mutate_caller_tools_list(chat_client_base: SupportsChatGetResponse):
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(_pte_factorial)
+        return "loaded"
+
+    original_tools: list[Any] = [load_math]
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": original_tools},
+    )
+    assert original_tools == [load_math]
+
+
+async def test_add_tools_persists_across_iterations(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    recorded: list[list[str]] = []
+    client_cls = type(chat_client_base)
+    original = client_cls._get_non_streaming_response
+
+    async def recording(self: Any, *, messages: Any, options: dict[str, Any], **kwargs: Any) -> ChatResponse:
+        tools = options.get("tools") or []
+        recorded.append([t.name for t in tools if isinstance(t, FunctionTool)])
+        return await original(self, messages=messages, options=options, **kwargs)
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(_pte_factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 4  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_function_call_response("3", "factorial", '{"n": 3}'),
+        _pte_text_response(),
+    ]
+    monkey = pytest.MonkeyPatch()
+    monkey.setattr(client_cls, "_get_non_streaming_response", recording)
+    try:
+        await chat_client_base.get_response(
+            [Message(role="user", contents=["hi"])],
+            options={"tool_choice": "auto", "tools": [load_math]},
+        )
+    finally:
+        monkey.undo()
+
+    assert "factorial" in recorded[1]
+    assert "factorial" in recorded[2]
+
+
+async def test_add_tools_through_function_middleware(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    class PassthroughMiddleware(FunctionMiddleware):
+        async def process(self, context: FunctionInvocationContext, call_next: Any) -> None:
+            await call_next()
+
+    @tool(name="factorial", approval_mode="never_require")
+    def factorial(n: int) -> int:
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [load_math]},
+        middleware=[PassthroughMiddleware()],
+    )
+    assert exec_counter == 1
+
+
+async def test_add_tools_with_approval_required_tool(chat_client_base: SupportsChatGetResponse):
+    @tool(name="secure_tool", approval_mode="always_require")
+    def secure_tool(value: str) -> str:
+        return f"secure: {value}"
+
+    @tool(name="load_secure", approval_mode="never_require")
+    def load_secure(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(secure_tool)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_secure"),
+        _pte_function_call_response("2", "secure_tool", '{"value": "x"}'),
+        _pte_text_response(),
+    ]
+    response = await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [load_secure]},
+    )
+    assert any(item.type == "function_approval_request" for msg in response.messages for item in msg.contents)
+
+
+async def test_add_tools_accepts_plain_callable(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    def plain_factorial(n: int) -> int:
+        """Compute factorial."""
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(plain_factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "plain_factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [load_math]},
+    )
+    assert exec_counter == 1
+
+
+async def test_add_tools_streaming(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    @tool(name="factorial", approval_mode="never_require")
+    def factorial(n: int) -> int:
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.streaming_responses = [
+        [
+            ChatResponseUpdate(
+                contents=[Content.from_function_call(call_id="1", name="load_math", arguments="{}")],
+                role="assistant",
+            )
+        ],
+        [
+            ChatResponseUpdate(
+                contents=[Content.from_function_call(call_id="2", name="factorial", arguments='{"n": 5}')],
+                role="assistant",
+            )
+        ],
+        [ChatResponseUpdate(contents=[Content.from_text("done")], role="assistant", finish_reason="stop")],
+    ]
+    async for _ in chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        stream=True,
+        options={"tool_choice": "auto", "tools": [load_math]},
+    ):
+        pass
+    assert exec_counter == 1
+
+
+def test_add_tools_duplicate_same_object_is_noop():
+    @tool(name="dup", approval_mode="never_require")
+    def dup(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=dup, arguments={}, tools=[dup])
+    ctx.add_tools(dup)
+    assert ctx.tools is not None
+    assert len(ctx.tools) == 1
+
+
+def test_add_tools_duplicate_name_different_object_raises():
+    @tool(name="dup", approval_mode="never_require")
+    def dup_a(x: int) -> int:
+        return x
+
+    @tool(name="dup", approval_mode="never_require")
+    def dup_b(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=dup_a, arguments={}, tools=[dup_a])
+    with pytest.raises(ValueError):
+        ctx.add_tools(dup_b)
+
+
+def test_add_tools_batch_with_duplicate_is_atomic():
+    """A duplicate-name clash partway through a batch must leave the live list unchanged."""
+
+    @tool(name="existing", approval_mode="never_require")
+    def existing(x: int) -> int:
+        return x
+
+    @tool(name="fresh", approval_mode="never_require")
+    def fresh(x: int) -> int:
+        return x
+
+    @tool(name="existing", approval_mode="never_require")
+    def clashing(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=existing, arguments={}, tools=[existing])
+    with pytest.raises(ValueError):
+        ctx.add_tools([fresh, clashing])
+    assert ctx.tools is not None
+    # The valid "fresh" tool must not have been committed before the clash raised.
+    assert ctx.tools == [existing]
+
+
+def test_remove_tools_by_name_and_object():
+    @tool(name="a", approval_mode="never_require")
+    def a(x: int) -> int:
+        return x
+
+    @tool(name="b", approval_mode="never_require")
+    def b(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=a, arguments={}, tools=[a, b])
+    ctx.remove_tools("a")
+    assert ctx.tools is not None
+    assert [t.name for t in ctx.tools] == ["b"]
+    ctx.remove_tools(b)
+    assert ctx.tools == []
+
+
+def test_remove_tools_unknown_name_is_noop():
+    @tool(name="a", approval_mode="never_require")
+    def a(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=a, arguments={}, tools=[a])
+    ctx.remove_tools("nonexistent")
+    assert ctx.tools is not None
+    assert [t.name for t in ctx.tools] == ["a"]
+
+
+def test_progressive_tools_helpers_raise_without_live_tools():
+    @tool(name="a", approval_mode="never_require")
+    def a(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=a, arguments={})
+    assert ctx.tools is None
+    with pytest.raises(RuntimeError):
+        ctx.add_tools(a)
+    with pytest.raises(RuntimeError):
+        ctx.remove_tools("a")
+
+
+# endregion

python/packages/core/tests/core/test_harness_agent.py

@@ -394,3 +394,94 @@ def test_create_harness_agent_logs_warning_when_no_web_search(caplog: pytest.Log
             max_output_tokens=16_384,
         )
     assert any("SupportsWebSearchTool" in msg for msg in caplog.messages)
+
+
+# --- Background Agents Tests ---
+
+
+class _FakeBackgroundAgent:
+    """Minimal agent stub satisfying SupportsAgentRun for background agents tests."""
+
+    def __init__(self, name: str, description: str | None = None):
+        self.id = f"agent-{name}"
+        self.name = name
+        self.description = description
+
+    def create_session(self, *, session_id: str | None = None) -> AgentSession:
+        return AgentSession(session_id=session_id)
+
+    def get_session(self, service_session_id: str, *, session_id: str | None = None) -> AgentSession:
+        return AgentSession(service_session_id=service_session_id, session_id=session_id)
+
+    async def run(self, messages: Any = None, *, stream: bool = False, session: Any = None, **kwargs: Any) -> Any:
+        from agent_framework import AgentResponse
+
+        return AgentResponse(messages=[], response_id="fake-bg-response")
+
+
+def test_create_harness_agent_no_background_agents_by_default() -> None:
+    """No BackgroundAgentsProvider should be included when background_agents is not provided."""
+    from agent_framework._harness._background_agents import BackgroundAgentsProvider
+
+    agent = create_harness_agent(
+        client=_FakeChatClient(),  # type: ignore[arg-type]
+        max_context_window_tokens=128_000,
+        max_output_tokens=16_384,
+        disable_web_search=True,
+    )
+    providers = agent.context_providers or []
+    assert not any(isinstance(p, BackgroundAgentsProvider) for p in providers)
+
+
+def test_create_harness_agent_adds_background_agents_provider() -> None:
+    """BackgroundAgentsProvider should be included when background_agents are provided."""
+    from agent_framework._harness._background_agents import BackgroundAgentsProvider
+
+    bg_agent = _FakeBackgroundAgent("WebSearcher", "Searches the web")
+    agent = create_harness_agent(
+        client=_FakeChatClient(),  # type: ignore[arg-type]
+        max_context_window_tokens=128_000,
+        max_output_tokens=16_384,
+        disable_web_search=True,
+        background_agents=[bg_agent],
+    )
+    providers = agent.context_providers or []
+    bg_providers = [p for p in providers if isinstance(p, BackgroundAgentsProvider)]
+    assert len(bg_providers) == 1
+
+
+def test_create_harness_agent_background_agents_custom_instructions() -> None:
+    """Custom instructions should be passed to BackgroundAgentsProvider."""
+    from agent_framework._harness._background_agents import BackgroundAgentsProvider
+
+    custom_instructions = "## Custom\n\nUse agents wisely.\n\n{background_agents}"
+    bg_agent = _FakeBackgroundAgent("Helper", "A helper agent")
+    agent = create_harness_agent(
+        client=_FakeChatClient(),  # type: ignore[arg-type]
+        max_context_window_tokens=128_000,
+        max_output_tokens=16_384,
+        disable_web_search=True,
+        background_agents=[bg_agent],
+        background_agents_instructions=custom_instructions,
+    )
+    providers = agent.context_providers or []
+    bg_providers = [p for p in providers if isinstance(p, BackgroundAgentsProvider)]
+    assert len(bg_providers) == 1
+    # Verify the custom instructions were used (placeholder replaced with agent list).
+    assert "Custom" in bg_providers[0]._instructions
+    assert "Helper" in bg_providers[0]._instructions
+
+
+def test_create_harness_agent_empty_background_agents_list() -> None:
+    """An empty background_agents list should NOT add a BackgroundAgentsProvider."""
+    from agent_framework._harness._background_agents import BackgroundAgentsProvider
+
+    agent = create_harness_agent(
+        client=_FakeChatClient(),  # type: ignore[arg-type]
+        max_context_window_tokens=128_000,
+        max_output_tokens=16_384,
+        disable_web_search=True,
+        background_agents=[],
+    )
+    providers = agent.context_providers or []
+    assert not any(isinstance(p, BackgroundAgentsProvider) for p in providers)

python/packages/core/tests/core/test_local_eval.py

@@ -11,8 +11,13 @@ import pytest
 from agent_framework._evaluation import (
     CheckResult,
     EvalItem,
+    EvalItemResult,
+    EvalNotPassedError,
+    EvalResults,
+    EvalScoreResult,
     ExpectedToolCall,
     LocalEvaluator,
+    RubricScore,
     _coerce_result,
     evaluator,
     keyword_check,
@@ -1010,19 +1015,101 @@ class TestAllPassedSubResults:
 
 
 # ---------------------------------------------------------------------------
-# r5 review: _build_overall_item with empty outputs
+# Rubric assertions (EvalResults.assert_*)
 # ---------------------------------------------------------------------------
 
 
-class TestBuildOverallItemEmpty:
-    """Test _build_overall_item returns None for empty workflow outputs."""
+def _rubric_results(*scores_per_item: list[EvalScoreResult]) -> EvalResults:
+    items = [
+        EvalItemResult(item_id=f"item-{i}", status="pass", scores=scores) for i, scores in enumerate(scores_per_item)
+    ]
+    return EvalResults(
+        provider="test",
+        eval_id="ev1",
+        run_id="run1",
+        result_counts={"passed": len(items), "failed": 0, "errored": 0, "total": len(items)},
+        items=items,
+    )
 
-    def test_returns_none_for_empty_outputs(self):
-        from unittest.mock import MagicMock
 
-        from agent_framework._evaluation import _build_overall_item
+class TestRubricAssertions:
+    """Tests for EvalResults.assert_dimension_score_at_least."""
 
-        mock_result = MagicMock()
-        mock_result.get_outputs.return_value = []
-        item = _build_overall_item("Hello", mock_result)
-        assert item is None
+    def test_dimension_at_or_above_threshold_passes(self) -> None:
+        results = _rubric_results(
+            [
+                EvalScoreResult(
+                    name="policy",
+                    score=0.9,
+                    dimensions=[RubricScore(id="clarity", score=4, applicable=True, weight=1, reason="")],
+                )
+            ],
+        )
+        # Should not raise.
+        results.assert_dimension_score_at_least("clarity", 3)
+
+    def test_dimension_below_threshold_raises(self) -> None:
+        results = _rubric_results(
+            [
+                EvalScoreResult(
+                    name="policy",
+                    score=0.5,
+                    dimensions=[RubricScore(id="clarity", score=2, applicable=True, weight=1, reason="")],
+                )
+            ],
+        )
+        with pytest.raises(EvalNotPassedError):
+            results.assert_dimension_score_at_least("clarity", 3)
+
+    def test_non_applicable_skipped_by_default(self) -> None:
+        results = _rubric_results(
+            [
+                EvalScoreResult(
+                    name="policy",
+                    score=1.0,
+                    dimensions=[RubricScore(id="clarity", score=None, applicable=False, weight=1, reason="n/a")],
+                )
+            ],
+        )
+        # No applicable scores; default behaviour is to skip silently.
+        results.assert_dimension_score_at_least("clarity", 3)
+
+    def test_require_applicable_raises_when_dimension_absent(self) -> None:
+        results = _rubric_results(
+            [EvalScoreResult(name="policy", score=1.0, dimensions=[])],
+        )
+        with pytest.raises(EvalNotPassedError, match="not applicable"):
+            results.assert_dimension_score_at_least("clarity", 3, require_applicable=True)
+
+    def test_require_applicable_raises_when_filtered_evaluator_missing(self) -> None:
+        # Regression: previously the (not evaluator or found_any) guard caused
+        # this case to silently pass even with require_applicable=True.
+        results = _rubric_results(
+            [
+                EvalScoreResult(
+                    name="other",
+                    score=0.9,
+                    dimensions=[RubricScore(id="clarity", score=4, applicable=True, weight=1, reason="")],
+                )
+            ],
+        )
+        with pytest.raises(EvalNotPassedError, match="not applicable"):
+            results.assert_dimension_score_at_least("clarity", 3, evaluator="policy", require_applicable=True)
+
+    def test_evaluator_filter_isolates_offenders(self) -> None:
+        results = _rubric_results(
+            [
+                EvalScoreResult(
+                    name="other",
+                    score=0.1,
+                    dimensions=[RubricScore(id="clarity", score=1, applicable=True, weight=1, reason="")],
+                ),
+                EvalScoreResult(
+                    name="policy",
+                    score=0.9,
+                    dimensions=[RubricScore(id="clarity", score=4, applicable=True, weight=1, reason="")],
+                ),
+            ],
+        )
+        # The low-scoring "other" evaluator is filtered out; "policy" passes.
+        results.assert_dimension_score_at_least("clarity", 3, evaluator="policy")

python/packages/core/tests/core/test_mcp_skills.py

@@ -0,0 +1,667 @@
+# Copyright (c) Microsoft. All rights reserved.
+
+"""Tests for MCP-based skills (MCPSkillsSource, MCPSkill, MCPSkillResource)."""
+
+from __future__ import annotations
+
+import base64
+import json
+from unittest.mock import AsyncMock
+
+import pytest
+from mcp.shared.exceptions import McpError
+from mcp.types import (
+    BlobResourceContents,
+    ErrorData,
+    ReadResourceResult,
+    TextResourceContents,
+)
+from pydantic import AnyUrl
+
+from agent_framework import MCPSkill, MCPSkillResource, MCPSkillsSource
+from agent_framework._skills import _parse_mcp_skill_index
+
+# ---------------------------------------------------------------------------
+# Fixtures & helpers
+# ---------------------------------------------------------------------------
+
+SAMPLE_SKILL_MD = """\
+---
+name: unit-converter
+description: Convert between common units.
+---
+# Unit Converter
+
+Body content here.
+"""
+
+SAMPLE_SKILL_INDEX = json.dumps(
+    {
+        "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
+        "skills": [
+            {
+                "name": "unit-converter",
+                "type": "skill-md",
+                "description": "Convert between common units.",
+                "url": "skill://unit-converter/SKILL.md",
+            }
+        ],
+    }
+)
+
+
+def _make_text_result(text: str, uri: str = "skill://test") -> ReadResourceResult:
+    """Create a ReadResourceResult with a single TextResourceContents."""
+    return ReadResourceResult(
+        contents=[TextResourceContents(uri=AnyUrl(uri), text=text, mimeType="text/markdown")]
+    )
+
+
+def _make_blob_result(
+    data: bytes,
+    uri: str = "skill://test",
+    mime_type: str = "application/octet-stream",
+) -> ReadResourceResult:
+    """Create a ReadResourceResult with a single BlobResourceContents."""
+    return ReadResourceResult(
+        contents=[BlobResourceContents(uri=AnyUrl(uri), blob=base64.b64encode(data).decode(), mimeType=mime_type)]
+    )
+
+
+def _make_empty_result() -> ReadResourceResult:
+    """Create a ReadResourceResult with no contents."""
+    return ReadResourceResult(contents=[])
+
+
+def _make_client(**read_resource_responses: ReadResourceResult) -> AsyncMock:
+    """Create a mock ClientSession whose read_resource returns different results per URI.
+
+    Args:
+        **read_resource_responses: Mapping of URI string to ReadResourceResult.
+            Any URI not in this mapping raises McpError with the MCP-spec
+            "Resource not found" code (-32002).
+    """
+    client = AsyncMock()
+
+    async def _read_resource(uri: AnyUrl) -> ReadResourceResult:
+        uri_str = str(uri)
+        if uri_str in read_resource_responses:
+            return read_resource_responses[uri_str]
+        raise McpError(error=ErrorData(code=-32002, message=f"Resource not found: {uri_str}"))
+
+    client.read_resource = AsyncMock(side_effect=_read_resource)
+    return client
+
+
+# ---------------------------------------------------------------------------
+# _parse_mcp_skill_index tests
+# ---------------------------------------------------------------------------
+
+
+class TestParseMCPSkillIndex:
+    """Tests for the _parse_mcp_skill_index helper."""
+
+    def test_parses_valid_index(self) -> None:
+        index = _parse_mcp_skill_index(SAMPLE_SKILL_INDEX)
+        assert index.schema == "https://schemas.agentskills.io/discovery/0.2.0/schema.json"
+        assert len(index.skills) == 1
+        assert index.skills[0].name == "unit-converter"
+        assert index.skills[0].type == "skill-md"
+        assert index.skills[0].url == "skill://unit-converter/SKILL.md"
+
+    def test_parses_empty_skills_array(self) -> None:
+        index = _parse_mcp_skill_index('{"$schema": "test", "skills": []}')
+        assert index.skills == []
+
+    def test_parses_missing_skills_key(self) -> None:
+        index = _parse_mcp_skill_index('{"$schema": "test"}')
+        assert index.skills == []
+
+    def test_raises_on_non_object(self) -> None:
+        with pytest.raises(ValueError, match="must be a JSON object"):
+            _parse_mcp_skill_index("[]")
+
+    def test_raises_on_invalid_json(self) -> None:
+        with pytest.raises(json.JSONDecodeError):
+            _parse_mcp_skill_index("not json")
+
+    def test_skips_non_dict_entries(self) -> None:
+        index = _parse_mcp_skill_index('{"skills": ["not-a-dict", {"name": "ok", "type": "skill-md"}]}')
+        assert len(index.skills) == 1
+        assert index.skills[0].name == "ok"
+
+
+# ---------------------------------------------------------------------------
+# MCPSkillResource tests
+# ---------------------------------------------------------------------------
+
+
+class TestMCPSkillResource:
+    """Tests for MCPSkillResource."""
+
+    @pytest.mark.asyncio
+    async def test_read_text_content(self) -> None:
+        result = _make_text_result("hello world")
+        resource = MCPSkillResource(name="test.md", result=result)
+        content = await resource.read()
+        assert content == "hello world"
+
+    @pytest.mark.asyncio
+    async def test_read_binary_content(self) -> None:
+        data = bytes([0x01, 0x02, 0x03, 0x04])
+        result = _make_blob_result(data)
+        resource = MCPSkillResource(name="icon.bin", result=result)
+        content = await resource.read()
+        assert content == data
+
+    @pytest.mark.asyncio
+    async def test_read_empty_returns_none(self) -> None:
+        result = _make_empty_result()
+        resource = MCPSkillResource(name="empty", result=result)
+        content = await resource.read()
+        assert content is None
+
+    @pytest.mark.asyncio
+    async def test_read_multiple_text_contents_joined(self) -> None:
+        result = ReadResourceResult(
+            contents=[
+                TextResourceContents(uri=AnyUrl("skill://a"), text="line1", mimeType="text/plain"),
+                TextResourceContents(uri=AnyUrl("skill://b"), text="line2", mimeType="text/plain"),
+            ]
+        )
+        resource = MCPSkillResource(name="multi", result=result)
+        content = await resource.read()
+        assert content == "line1\nline2"
+
+    @pytest.mark.asyncio
+    async def test_binary_takes_precedence_over_text(self) -> None:
+        data = b"\xff\xfe"
+        result = ReadResourceResult(
+            contents=[
+                TextResourceContents(uri=AnyUrl("skill://a"), text="text", mimeType="text/plain"),
+                BlobResourceContents(
+                    uri=AnyUrl("skill://b"),
+                    blob=base64.b64encode(data).decode(),
+                    mimeType="application/octet-stream",
+                ),
+            ]
+        )
+        resource = MCPSkillResource(name="mixed", result=result)
+        content = await resource.read()
+        # The implementation iterates all contents checking for BlobResourceContents
+        # first, so when both text and binary are present, binary is returned.
+        assert content == data
+
+
+# ---------------------------------------------------------------------------
+# MCPSkill tests
+# ---------------------------------------------------------------------------
+
+
+class TestMCPSkill:
+    """Tests for MCPSkill."""
+
+    @pytest.mark.asyncio
+    async def test_get_content_fetches_and_caches(self) -> None:
+        client = _make_client(**{"skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD)})
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="unit-converter", description="Convert between common units.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://unit-converter/SKILL.md", client=client)
+
+        content1 = await skill.get_content()
+        content2 = await skill.get_content()
+
+        assert "Body content here." in content1
+        assert content1 == content2
+        # Only one MCP call should be made (cached)
+        assert client.read_resource.call_count == 1
+
+    @pytest.mark.asyncio
+    async def test_get_content_raises_on_empty(self) -> None:
+        client = _make_client(**{"skill://empty/SKILL.md": _make_empty_result()})
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="empty-skill", description="Empty skill.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://empty/SKILL.md", client=client)
+
+        with pytest.raises(ValueError, match="no text content"):
+            await skill.get_content()
+
+    @pytest.mark.asyncio
+    async def test_get_resource_text(self) -> None:
+        client = _make_client(
+            **{
+                "skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD),
+                "skill://unit-converter/references/checklist.md": _make_text_result("- check thing 1\n- check thing 2"),
+            }
+        )
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="unit-converter", description="Convert between common units.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://unit-converter/SKILL.md", client=client)
+
+        resource = await skill.get_resource("references/checklist.md")
+        assert resource is not None
+        content = await resource.read()
+        assert content == "- check thing 1\n- check thing 2"
+
+    @pytest.mark.asyncio
+    async def test_get_resource_binary(self) -> None:
+        data = bytes([0x01, 0x02, 0x03, 0x04])
+        client = _make_client(
+            **{
+                "skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD),
+                "skill://unit-converter/assets/icon.bin": _make_blob_result(data),
+            }
+        )
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="unit-converter", description="Convert between common units.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://unit-converter/SKILL.md", client=client)
+
+        resource = await skill.get_resource("assets/icon.bin")
+        assert resource is not None
+        content = await resource.read()
+        assert content == data
+
+    @pytest.mark.asyncio
+    async def test_get_resource_unknown_returns_none(self) -> None:
+        client = _make_client(**{"skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD)})
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="unit-converter", description="Convert between common units.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://unit-converter/SKILL.md", client=client)
+
+        resource = await skill.get_resource("references/does-not-exist.md")
+        assert resource is None
+
+    @pytest.mark.asyncio
+    @pytest.mark.parametrize(
+        "name",
+        [
+            "../escape.md",
+            "references/../../escape.md",
+            "..",
+            "..\\escape.md",
+            "/etc/passwd",
+            "http://attacker.example.com/payload",
+        ],
+    )
+    async def test_get_resource_path_traversal_returns_none(self, name: str) -> None:
+        # Register a permissive mock that would happily return content for any URI,
+        # so the test fails unless the client-side validation rejects the name
+        # before issuing the read.
+        client = AsyncMock()
+        client.read_resource = AsyncMock(return_value=_make_text_result("should never be returned"))
+
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="unit-converter", description="Convert between common units.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://unit-converter/SKILL.md", client=client)
+
+        resource = await skill.get_resource(name)
+        assert resource is None
+        client.read_resource.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_get_resource_empty_name_returns_none(self) -> None:
+        client = _make_client()
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+
+        assert await skill.get_resource("") is None
+        assert await skill.get_resource("   ") is None
+
+    @pytest.mark.asyncio
+    async def test_get_script_returns_none(self) -> None:
+        client = _make_client()
+        from agent_framework import SkillFrontmatter
+
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+
+        assert await skill.get_script("anything") is None
+
+    def test_compute_skill_root_uri_strips_suffix(self) -> None:
+        assert MCPSkill._compute_skill_root_uri("skill://unit-converter/SKILL.md") == "skill://unit-converter/"
+
+    def test_compute_skill_root_uri_trailing_slash(self) -> None:
+        assert MCPSkill._compute_skill_root_uri("skill://unit-converter/") == "skill://unit-converter/"
+
+    def test_compute_skill_root_uri_no_suffix_adds_slash(self) -> None:
+        assert MCPSkill._compute_skill_root_uri("skill://unit-converter") == "skill://unit-converter/"
+
+
+# ---------------------------------------------------------------------------
+# MCPSkillsSource tests
+# ---------------------------------------------------------------------------
+
+
+class TestMCPSkillsSource:
+    """Tests for MCPSkillsSource."""
+
+    @pytest.mark.asyncio
+    async def test_index_based_discovery_returns_skill(self) -> None:
+        client = _make_client(
+            **{
+                "skill://index.json": _make_text_result(SAMPLE_SKILL_INDEX, uri="skill://index.json"),
+                "skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD),
+            }
+        )
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+
+        assert len(skills) == 1
+        assert skills[0].frontmatter.name == "unit-converter"
+        assert skills[0].frontmatter.description == "Convert between common units."
+
+        # Content is fetched on demand, not during discovery
+        content = await skills[0].get_content()
+        assert "Body content here." in content
+
+    @pytest.mark.asyncio
+    async def test_no_index_returns_empty(self) -> None:
+        client = _make_client()  # No resources at all
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_does_not_read_skill_md_during_discovery(self) -> None:
+        # Index points to a skill, but SKILL.md is not registered on the server.
+        # Discovery should succeed because it only reads the index.
+        client = _make_client(
+            **{"skill://index.json": _make_text_result(SAMPLE_SKILL_INDEX, uri="skill://index.json")}
+        )
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+
+        assert len(skills) == 1
+        assert skills[0].frontmatter.name == "unit-converter"
+
+    @pytest.mark.asyncio
+    async def test_invalid_name_is_skipped(self) -> None:
+        index_json = json.dumps(
+            {
+                "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
+                "skills": [
+                    {
+                        "name": "UnitConverter",  # Invalid: uppercase
+                        "type": "skill-md",
+                        "description": "Convert between common units.",
+                        "url": "skill://UnitConverter/SKILL.md",
+                    }
+                ],
+            }
+        )
+        client = _make_client(**{"skill://index.json": _make_text_result(index_json, uri="skill://index.json")})
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_missing_required_fields_is_skipped(self) -> None:
+        index_json = json.dumps(
+            {
+                "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
+                "skills": [
+                    {
+                        "name": "unit-converter",
+                        "type": "skill-md",
+                        # Missing description and url
+                    }
+                ],
+            }
+        )
+        client = _make_client(**{"skill://index.json": _make_text_result(index_json, uri="skill://index.json")})
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_unsupported_type_is_skipped(self) -> None:
+        index_json = json.dumps(
+            {
+                "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
+                "skills": [
+                    {
+                        "name": "some-skill",
+                        "type": "archive",
+                        "description": "Packaged skill.",
+                        "url": "skill://some-skill.tar.gz",
+                    }
+                ],
+            }
+        )
+        client = _make_client(**{"skill://index.json": _make_text_result(index_json, uri="skill://index.json")})
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_template_type_is_skipped(self) -> None:
+        index_json = json.dumps(
+            {
+                "$schema": "https://schemas.agentskills.io/discovery/0.2.0/schema.json",
+                "skills": [
+                    {
+                        "type": "mcp-resource-template",
+                        "description": "Per-product documentation skill",
+                        "url": "skill://docs/{product}/SKILL.md",
+                    }
+                ],
+            }
+        )
+        client = _make_client(**{"skill://index.json": _make_text_result(index_json, uri="skill://index.json")})
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_empty_index_returns_empty(self) -> None:
+        client = _make_client(
+            **{"skill://index.json": _make_text_result('{"skills": []}', uri="skill://index.json")}
+        )
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_malformed_index_json_returns_empty(self) -> None:
+        client = _make_client(
+            **{"skill://index.json": _make_text_result("not valid json", uri="skill://index.json")}
+        )
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_sibling_text_resource(self) -> None:
+        client = _make_client(
+            **{
+                "skill://index.json": _make_text_result(SAMPLE_SKILL_INDEX, uri="skill://index.json"),
+                "skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD),
+                "skill://unit-converter/references/checklist.md": _make_text_result("- check thing 1\n- check thing 2"),
+            }
+        )
+        source = MCPSkillsSource(client=client)
+        skill = (await source.get_skills())[0]
+        resource = await skill.get_resource("references/checklist.md")
+        assert resource is not None
+        content = await resource.read()
+        assert content == "- check thing 1\n- check thing 2"
+
+    @pytest.mark.asyncio
+    async def test_sibling_binary_resource(self) -> None:
+        data = bytes([0x01, 0x02, 0x03, 0x04])
+        client = _make_client(
+            **{
+                "skill://index.json": _make_text_result(SAMPLE_SKILL_INDEX, uri="skill://index.json"),
+                "skill://unit-converter/SKILL.md": _make_text_result(SAMPLE_SKILL_MD),
+                "skill://unit-converter/assets/icon.bin": _make_blob_result(data),
+            }
+        )
+        source = MCPSkillsSource(client=client)
+        skill = (await source.get_skills())[0]
+        resource = await skill.get_resource("assets/icon.bin")
+        assert resource is not None
+        content = await resource.read()
+        assert content == data
+
+
+# ---------------------------------------------------------------------------
+# McpError code branching tests
+# ---------------------------------------------------------------------------
+
+
+class TestMCPSkillsSourceErrorCodeBranching:
+    """Tests that MCPSkillsSource and MCPSkill branch on McpError.error.code.
+
+    Only "not found" codes (RESOURCE_NOT_FOUND -32002, METHOD_NOT_FOUND -32601)
+    should be silently swallowed as "no skills available." Other McpError codes
+    and non-McpError exceptions must propagate so that auth failures, server
+    crashes, and connection drops are visible.
+    """
+
+    @pytest.mark.asyncio
+    async def test_index_method_not_found_returns_empty(self) -> None:
+        """METHOD_NOT_FOUND (-32601) -> server doesn't support resources/read."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=McpError(error=ErrorData(code=-32601, message="Method not found")))
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_index_resource_not_found_returns_empty(self) -> None:
+        """MCP-spec "Resource not found" (-32002) -> server has no index."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(
+            side_effect=McpError(error=ErrorData(code=-32002, message="Resource not found"))
+        )
+        source = MCPSkillsSource(client=client)
+        skills = await source.get_skills()
+        assert skills == []
+
+    @pytest.mark.asyncio
+    async def test_index_invalid_params_propagates(self) -> None:
+        """INVALID_PARAMS (-32602) is a real bug, must propagate (not "not found")."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=McpError(error=ErrorData(code=-32602, message="Invalid params")))
+        source = MCPSkillsSource(client=client)
+        with pytest.raises(McpError):
+            await source.get_skills()
+
+    @pytest.mark.asyncio
+    async def test_index_internal_error_propagates(self) -> None:
+        """INTERNAL_ERROR (-32603) must propagate, not silently return empty."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=McpError(error=ErrorData(code=-32603, message="Internal error")))
+        source = MCPSkillsSource(client=client)
+        with pytest.raises(McpError):
+            await source.get_skills()
+
+    @pytest.mark.asyncio
+    async def test_index_connection_closed_propagates(self) -> None:
+        """CONNECTION_CLOSED (-32000) must propagate."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(
+            side_effect=McpError(error=ErrorData(code=-32000, message="Connection closed"))
+        )
+        source = MCPSkillsSource(client=client)
+        with pytest.raises(McpError):
+            await source.get_skills()
+
+    @pytest.mark.asyncio
+    async def test_index_generic_error_code_propagates(self) -> None:
+        """Generic handler error (code 0) must propagate."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=McpError(error=ErrorData(code=0, message="Some handler error")))
+        source = MCPSkillsSource(client=client)
+        with pytest.raises(McpError):
+            await source.get_skills()
+
+    @pytest.mark.asyncio
+    async def test_index_non_mcp_error_propagates(self) -> None:
+        """Non-McpError exceptions (connection drop, timeout) must propagate."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=ConnectionError("connection lost"))
+        source = MCPSkillsSource(client=client)
+        with pytest.raises(ConnectionError):
+            await source.get_skills()
+
+    @pytest.mark.asyncio
+    async def test_get_resource_internal_error_propagates(self) -> None:
+        """McpError with INTERNAL_ERROR on get_resource must propagate."""
+        from agent_framework import SkillFrontmatter
+
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=McpError(error=ErrorData(code=-32603, message="Server crashed")))
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+        with pytest.raises(McpError):
+            await skill.get_resource("references/file.md")
+
+    @pytest.mark.asyncio
+    async def test_get_resource_not_found_returns_none(self) -> None:
+        """McpError with RESOURCE_NOT_FOUND (-32002) on get_resource returns None."""
+        from agent_framework import SkillFrontmatter
+
+        client = AsyncMock()
+        client.read_resource = AsyncMock(
+            side_effect=McpError(error=ErrorData(code=-32002, message="Resource not found"))
+        )
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+        result = await skill.get_resource("references/file.md")
+        assert result is None
+
+    @pytest.mark.asyncio
+    async def test_get_resource_connection_error_propagates(self) -> None:
+        """A plain ConnectionError on get_resource must propagate, not return None."""
+        from agent_framework import SkillFrontmatter
+
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=ConnectionError("connection lost"))
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+        with pytest.raises(ConnectionError):
+            await skill.get_resource("references/file.md")
+
+    @pytest.mark.asyncio
+    async def test_get_resource_timeout_error_propagates(self) -> None:
+        """A TimeoutError on get_resource must propagate, not return None."""
+        from agent_framework import SkillFrontmatter
+
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=TimeoutError("read timed out"))
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+        with pytest.raises(TimeoutError):
+            await skill.get_resource("references/file.md")
+
+    @pytest.mark.asyncio
+    async def test_get_resource_generic_mcp_error_propagates(self) -> None:
+        """McpError with a generic code (0) on get_resource must propagate."""
+        from agent_framework import SkillFrontmatter
+
+        client = AsyncMock()
+        client.read_resource = AsyncMock(
+            side_effect=McpError(error=ErrorData(code=0, message="Handler error"))
+        )
+        fm = SkillFrontmatter(name="test-skill", description="Test.")
+        skill = MCPSkill(frontmatter=fm, skill_md_uri="skill://test/SKILL.md", client=client)
+        with pytest.raises(McpError):
+            await skill.get_resource("references/file.md")
+
+    @pytest.mark.asyncio
+    async def test_index_timeout_error_propagates(self) -> None:
+        """A TimeoutError reading skill://index.json must propagate."""
+        client = AsyncMock()
+        client.read_resource = AsyncMock(side_effect=TimeoutError("read timed out"))
+        source = MCPSkillsSource(client=client)
+        with pytest.raises(TimeoutError):
+            await source.get_skills()

python/packages/core/tests/core/test_observability.py

@@ -761,6 +761,115 @@ def test_get_exporters_from_env_missing_grpc_dependency(monkeypatch):
         _get_exporters_from_env()
 
 
+# region Test OTLP endpoint computation (base-URL auto-append for HTTP)
+
+
+def test_get_exporters_from_env_http_base_endpoint_appends_signal_paths(monkeypatch):
+    """OTEL_EXPORTER_OTLP_ENDPOINT is a base URL for HTTP; SDK auto-appends
+    /v1/{traces,metrics,logs}. Because we read the env var and forward it as the
+    constructor ``endpoint=`` arg (which the SDK treats as a full URL), we must
+    replicate the auto-append ourselves.
+    """
+    from unittest.mock import patch
+
+    from agent_framework import observability
+
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://localhost:4318")
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_PROTOCOL", "http/protobuf")
+    for key in (
+        "OTEL_EXPORTER_OTLP_TRACES_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_METRICS_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_LOGS_ENDPOINT",
+    ):
+        monkeypatch.delenv(key, raising=False)
+
+    with patch.object(observability, "_create_otlp_exporters", return_value=[]) as create:
+        observability._get_exporters_from_env()
+
+    kwargs = create.call_args.kwargs
+    assert kwargs["protocol"] == "http/protobuf"
+    assert kwargs["traces_endpoint"] == "http://localhost:4318/v1/traces"
+    assert kwargs["metrics_endpoint"] == "http://localhost:4318/v1/metrics"
+    assert kwargs["logs_endpoint"] == "http://localhost:4318/v1/logs"
+
+
+def test_get_exporters_from_env_http_base_endpoint_trailing_slash(monkeypatch):
+    """A trailing slash on the base endpoint should not produce a doubled slash."""
+    from unittest.mock import patch
+
+    from agent_framework import observability
+
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://localhost:4318/")
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_PROTOCOL", "http/protobuf")
+    for key in (
+        "OTEL_EXPORTER_OTLP_TRACES_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_METRICS_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_LOGS_ENDPOINT",
+    ):
+        monkeypatch.delenv(key, raising=False)
+
+    with patch.object(observability, "_create_otlp_exporters", return_value=[]) as create:
+        observability._get_exporters_from_env()
+
+    kwargs = create.call_args.kwargs
+    assert kwargs["traces_endpoint"] == "http://localhost:4318/v1/traces"
+    assert kwargs["metrics_endpoint"] == "http://localhost:4318/v1/metrics"
+    assert kwargs["logs_endpoint"] == "http://localhost:4318/v1/logs"
+
+
+def test_get_exporters_from_env_http_signal_specific_used_verbatim(monkeypatch):
+    """Signal-specific endpoint env vars are full URLs and must be used verbatim,
+    even when a base endpoint is also set.
+    """
+    from unittest.mock import patch
+
+    from agent_framework import observability
+
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://localhost:4318")
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_TRACES_ENDPOINT", "http://traces.example.com/custom/path")
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_PROTOCOL", "http/protobuf")
+    for key in (
+        "OTEL_EXPORTER_OTLP_METRICS_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_LOGS_ENDPOINT",
+    ):
+        monkeypatch.delenv(key, raising=False)
+
+    with patch.object(observability, "_create_otlp_exporters", return_value=[]) as create:
+        observability._get_exporters_from_env()
+
+    kwargs = create.call_args.kwargs
+    # Signal-specific is verbatim — no path appended
+    assert kwargs["traces_endpoint"] == "http://traces.example.com/custom/path"
+    # Others fall back to base, with path appended
+    assert kwargs["metrics_endpoint"] == "http://localhost:4318/v1/metrics"
+    assert kwargs["logs_endpoint"] == "http://localhost:4318/v1/logs"
+
+
+def test_get_exporters_from_env_grpc_base_endpoint_unchanged(monkeypatch):
+    """For gRPC, the base endpoint applies to all signals as-is (no path append)."""
+    from unittest.mock import patch
+
+    from agent_framework import observability
+
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_ENDPOINT", "http://localhost:4317")
+    monkeypatch.setenv("OTEL_EXPORTER_OTLP_PROTOCOL", "grpc")
+    for key in (
+        "OTEL_EXPORTER_OTLP_TRACES_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_METRICS_ENDPOINT",
+        "OTEL_EXPORTER_OTLP_LOGS_ENDPOINT",
+    ):
+        monkeypatch.delenv(key, raising=False)
+
+    with patch.object(observability, "_create_otlp_exporters", return_value=[]) as create:
+        observability._get_exporters_from_env()
+
+    kwargs = create.call_args.kwargs
+    assert kwargs["protocol"] == "grpc"
+    assert kwargs["traces_endpoint"] == "http://localhost:4317"
+    assert kwargs["metrics_endpoint"] == "http://localhost:4317"
+    assert kwargs["logs_endpoint"] == "http://localhost:4317"
+
+
 # region Test create_resource
 
 
@@ -1691,6 +1800,65 @@ def test_to_otel_part_function_call():
     }
 
 
+def test_to_otel_part_function_call_reuses_prepared_arguments():
+    """Test _to_otel_part does not re-serialize function-call arguments in the observability hot path."""
+    from agent_framework import Content
+    from agent_framework.observability import _to_otel_part
+
+    arguments = {"payload": object()}
+    content = Content(type="function_call", call_id="call_789", name="handoff", arguments=arguments)
+    result = _to_otel_part(content)
+
+    assert result is not None
+    assert result["arguments"] is arguments
+
+
+def test_make_json_safe_non_callable_method_attribute():
+    """Test make_json_safe handles objects where model_dump/to_dict/dict are non-callable attributes."""
+    from agent_framework._serialization import make_json_safe
+
+    class ObjWithNonCallableModelDump:
+        model_dump = 42  # not callable
+
+    obj = ObjWithNonCallableModelDump()
+    result = make_json_safe(obj)
+    assert result == {}
+
+
+def test_make_json_safe_callable_method_type_error_falls_through():
+    """Test make_json_safe falls through when serializer-like methods require arguments."""
+    from agent_framework._serialization import make_json_safe
+
+    class ObjWithRequiredArgModelDump:
+        def __init__(self) -> None:
+            self.value = "fallback"
+
+        def model_dump(self, required: str) -> dict[str, str]:
+            return {"required": required}
+
+    obj = ObjWithRequiredArgModelDump()
+    result = make_json_safe(obj)
+    assert result == {"value": "fallback"}
+
+
+def test_make_json_safe_dict_with_non_string_keys():
+    """Test make_json_safe converts non-primitive dict keys to strings."""
+    import json
+    from datetime import datetime
+
+    from agent_framework._serialization import make_json_safe
+
+    dt_key = datetime(2024, 1, 1)
+    obj = {dt_key: "value", 42: "num_value", "str_key": "normal"}
+    result = make_json_safe(obj)
+    # json.dumps must not raise TypeError
+    serialized = json.dumps(result)
+    parsed = json.loads(serialized)
+    assert parsed[str(dt_key)] == "value"
+    assert parsed["42"] == "num_value"
+    assert parsed["str_key"] == "normal"
+
+
 def test_to_otel_part_function_result():
     """Test _to_otel_part with function_result content."""
     from agent_framework import Content
@@ -3019,6 +3187,49 @@ async def test_system_instructions_preserves_non_ascii_characters(span_exporter:
     assert [msg.get("role") for msg in input_messages] == ["user"]
 
 
+@pytest.mark.parametrize("enable_sensitive_data", [True], indirect=True)
+def test_capture_messages_with_prepared_request_info_function_call_arguments(span_exporter: InMemorySpanExporter):
+    """Test _capture_messages handles request-info function-call arguments prepared at Content creation."""
+    import dataclasses
+    import json
+
+    from opentelemetry import trace
+
+    from agent_framework import WorkflowAgent
+
+    @dataclasses.dataclass
+    class HandoffRequest:
+        target_agent: str
+        reason: str
+
+    arguments = WorkflowAgent.RequestInfoFunctionArgs(
+        request_id="call_dc",
+        data=HandoffRequest(target_agent="helper", reason="overflow"),
+    ).to_dict()
+    msg = Message(
+        role="assistant",
+        contents=[
+            Content(
+                type="function_call",
+                call_id="call_dc",
+                name="request_info",
+                arguments=arguments,
+            )
+        ],
+    )
+    span_exporter.clear()
+    tracer = trace.get_tracer("test")
+    with tracer.start_as_current_span("test_span") as span:
+        _capture_messages(span=span, provider_name="test_provider", messages=[msg])
+
+    spans = span_exporter.get_finished_spans()
+    span = spans[0]
+    input_messages = json.loads(span.attributes[OtelAttr.INPUT_MESSAGES])
+    tool_part = input_messages[0]["parts"][0]
+    assert tool_part["type"] == "tool_call"
+    assert tool_part["arguments"]["data"] == {"target_agent": "helper", "reason": "overflow"}
+
+
 def test_capture_messages_keeps_framework_instructions_out_of_logs_and_span_messages(
     span_exporter: InMemorySpanExporter,
 ):

python/packages/core/tests/core/test_skills.py

@@ -4086,8 +4086,8 @@ class TestClassSkill:
 
     async def test_content_is_cached(self) -> None:
         skill = _MinimalClassSkill()
-        content1 = (await skill.get_content())
-        content2 = (await skill.get_content())
+        content1 = await skill.get_content()
+        content2 = await skill.get_content()
         assert content1 is content2
 
     def test_resources_are_lazy_cached(self) -> None:
@@ -5587,8 +5587,8 @@ class TestInlineSkillContentCaching:
     async def test_content_cached_after_first_access(self) -> None:
         """InlineSkill.content returns the same object on subsequent accesses."""
         skill = InlineSkill(frontmatter=SkillFrontmatter(name="test-skill", description="Test"), instructions="Body")
-        first = (await skill.get_content())
-        second = (await skill.get_content())
+        first = await skill.get_content()
+        second = await skill.get_content()
         assert first is second  # Same object (cached)
         assert "<name>test-skill</name>" in first
 

python/packages/core/tests/workflow/test_functional_workflow.py

@@ -5,6 +5,7 @@
 from __future__ import annotations
 
 import asyncio
+import json
 import logging
 from collections.abc import Iterator
 from contextlib import contextmanager
@@ -1642,6 +1643,37 @@ class TestFunctionalWorkflowAgentHITL:
                     break
         assert approval_found, "expected FunctionApprovalRequestContent in agent response"
 
+    async def test_request_info_dataclass_arguments_are_serialized_for_agent(self):
+        @dataclass
+        class HandoffRequest:
+            target_agent: str
+            reason: str
+
+        @workflow
+        async def wf(x: str, ctx: RunContext) -> str:
+            answer = await ctx.request_info(
+                HandoffRequest(target_agent=x, reason="overflow"),
+                response_type=str,
+                request_id="rid-1",
+            )
+            return f"got:{answer}"
+
+        agent = wf.as_agent()
+        response = await agent.run("helper")
+
+        function_call_arguments = None
+        for message in response.messages:
+            for content in message.contents:
+                if getattr(content, "type", None) == "function_approval_request" and content.function_call is not None:
+                    function_call_arguments = content.function_call.arguments
+                    break
+
+        assert function_call_arguments == {
+            "request_id": "rid-1",
+            "data": {"target_agent": "helper", "reason": "overflow"},
+        }
+        assert json.loads(json.dumps(function_call_arguments)) == function_call_arguments
+
     async def test_resume_via_agent_responses_kwarg(self):
         @workflow
         async def wf(x: str, ctx: RunContext) -> str:

python/packages/core/tests/workflow/test_workflow_agent.py

@@ -1,7 +1,9 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+import json
 import uuid
 from collections.abc import Awaitable, Sequence
+from dataclasses import dataclass
 from typing import Any, Literal, overload
 
 import pytest
@@ -23,6 +25,7 @@ from agent_framework import (
     WorkflowAgent,
     WorkflowBuilder,
     WorkflowContext,
+    WorkflowEvent,
     executor,
     handler,
     response_handler,
@@ -293,6 +296,33 @@ class TestWorkflowAgent:
         # Verify cleanup - pending requests should be cleared after function response handling
         assert len(agent.pending_requests) == 0
 
+    def test_request_info_dataclass_arguments_are_serialized_when_content_is_created(self) -> None:
+        """Test WorkflowAgent prepares request_info arguments before observability captures messages."""
+
+        @dataclass
+        class HandoffRequest:
+            target_agent: str
+            reason: str
+
+        executor = SimpleExecutor(id="executor1", response_text="Response")
+        workflow = WorkflowBuilder(start_executor=executor).build()
+        agent = WorkflowAgent(workflow=workflow, name="Request Test Agent")
+        event = WorkflowEvent.request_info(
+            request_id="request_123",
+            source_executor_id="executor1",
+            request_data=HandoffRequest(target_agent="helper", reason="overflow"),
+            response_type=str,
+        )
+
+        function_call, approval_request = agent._process_request_info_event(event)  # pyright: ignore[reportPrivateUsage]
+
+        assert function_call.arguments == {
+            "request_id": "request_123",
+            "data": {"target_agent": "helper", "reason": "overflow"},
+        }
+        assert approval_request.function_call is function_call
+        assert json.loads(json.dumps(function_call.arguments)) == function_call.arguments
+
     def test_workflow_as_agent_method(self) -> None:
         """Test that Workflow.as_agent() creates a properly configured WorkflowAgent."""
         # Create a simple workflow

python/packages/declarative/README.md

@@ -6,6 +6,18 @@ Please install this package via pip:
 pip install agent-framework-declarative --pre
 ```
 
+## Release stage
+
+This package ships at two different stability levels:
+
+- **Declarative workflows** (`WorkflowFactory`, executors, handlers, and the
+  `_workflows` surface) are at **release-candidate** stability and may receive only
+  minor refinements before GA.
+- **Declarative agents** (`AgentFactory` and the YAML agent loading/parsing path:
+  `DeclarativeLoaderError`, `ProviderLookupError`, `ProviderTypeMapping`) are
+  **experimental** and may change or be removed in future versions without notice.
+  Using any of these symbols emits an `ExperimentalWarning` on first use.
+
 ## Declarative features
 
 The declarative packages provides support for building agents based on a declarative yaml specification.

python/packages/declarative/agent_framework_declarative/__init__.py

@@ -1,5 +1,18 @@
 # Copyright (c) Microsoft. All rights reserved.
 
+"""Declarative specification support for Microsoft Agent Framework.
+
+Release stage:
+
+* The declarative-workflows surface (``WorkflowFactory``, executors, handlers,
+  etc.) is at release-candidate stability.
+* The declarative-agents surface (``AgentFactory`` and the YAML agent
+  loading/parsing path: ``DeclarativeLoaderError``, ``ProviderLookupError``,
+  ``ProviderTypeMapping``) is *experimental* and may change or be removed in
+  future versions without notice. Using these symbols emits an
+  ``ExperimentalWarning`` on first use.
+"""
+
 from importlib import metadata
 
 from ._loader import AgentFactory, DeclarativeLoaderError, ProviderLookupError, ProviderTypeMapping

python/packages/declarative/agent_framework_declarative/_loader.py

@@ -15,6 +15,10 @@ from agent_framework import (
 from agent_framework import (
     FunctionTool as AFFunctionTool,
 )
+from agent_framework._feature_stage import (  # type: ignore[reportPrivateUsage]
+    ExperimentalFeature,
+    experimental,
+)
 from agent_framework.exceptions import AgentException
 from dotenv import load_dotenv
 
@@ -43,6 +47,7 @@ else:
     from typing_extensions import TypedDict  # type: ignore # pragma: no cover
 
 
+@experimental(feature_id=ExperimentalFeature.DECLARATIVE_AGENTS)
 class ProviderTypeMapping(TypedDict, total=True):
     package: str
     name: str
@@ -118,18 +123,21 @@ PROVIDER_TYPE_OBJECT_MAPPING: dict[str, ProviderTypeMapping] = {
 }
 
 
+@experimental(feature_id=ExperimentalFeature.DECLARATIVE_AGENTS)
 class DeclarativeLoaderError(AgentException):
     """Exception raised for errors in the declarative loader."""
 
     pass
 
 
+@experimental(feature_id=ExperimentalFeature.DECLARATIVE_AGENTS)
 class ProviderLookupError(DeclarativeLoaderError):
     """Exception raised for errors in provider type lookup."""
 
     pass
 
 
+@experimental(feature_id=ExperimentalFeature.DECLARATIVE_AGENTS)
 class AgentFactory:
     """Factory for creating Agent instances from declarative YAML definitions.
 

python/packages/declarative/pyproject.toml

@@ -4,7 +4,7 @@ description = "Declarative specification support for Microsoft Agent Framework."
 authors = [{ name = "Microsoft", email = "af-support@microsoft.com"}]
 readme = "README.md"
 requires-python = ">=3.10"
-version = "1.0.0b260528"
+version = "1.0.0rc1"
 license-files = ["LICENSE"]
 urls.homepage = "https://aka.ms/agent-framework"
 urls.source = "https://github.com/microsoft/agent-framework/tree/main/python"
@@ -29,7 +29,7 @@ dependencies = [
 ]
 [dependency-groups]
 dev = [
-    "types-PyYaml==6.0.12.20250915"
+    "types-PyYaml==6.0.12.20260518"
 ]
 
 [tool.uv]
@@ -49,7 +49,8 @@ addopts = "-ra -q -r fEX"
 asyncio_mode = "auto"
 asyncio_default_fixture_loop_scope = "function"
 filterwarnings = [
-    "ignore:Support for class-based `config` is deprecated:DeprecationWarning:pydantic.*"
+    "ignore:Support for class-based `config` is deprecated:DeprecationWarning:pydantic.*",
+    "ignore::agent_framework._feature_stage.ExperimentalWarning",
 ]
 timeout = 120
 markers = [

python/packages/devui/agent_framework_devui/_server.py

@@ -375,13 +375,15 @@ class DevServer:
             logger.info("Starting Agent Framework Server")
             await self._ensure_executor()
             await self._ensure_openai_executor()  # Initialize OpenAI executor
-            yield
-            # Shutdown
-            logger.info("Shutting down Agent Framework Server")
+            try:
+                yield
+            finally:
+                # Shutdown
+                logger.info("Shutting down Agent Framework Server")
 
-            # Cleanup entity resources (e.g., close credentials, clients)
-            if self.executor:
-                await self._cleanup_entities()
+                # Cleanup entity resources (e.g., close credentials, clients)
+                if self.executor:
+                    await self._cleanup_entities()
 
         app = FastAPI(
             title="Agent Framework Server",

python/packages/durabletask/pyproject.toml

@@ -30,7 +30,7 @@ dependencies = [
 
 [dependency-groups]
 dev = [
-    "types-python-dateutil==2.9.0.20260402",
+    "types-python-dateutil==2.9.0.20260518",
 ]
 
 [tool.uv]

python/packages/foundry/agent_framework_foundry/__init__.py

@@ -12,6 +12,7 @@ from ._embedding_client import (
 )
 from ._foundry_evals import (
     FoundryEvals,
+    GeneratedEvaluatorRef,
     evaluate_foundry_target,
     evaluate_traces,
 )
@@ -33,6 +34,7 @@ __all__ = [
     "FoundryEmbeddingSettings",
     "FoundryEvals",
     "FoundryMemoryProvider",
+    "GeneratedEvaluatorRef",
     "RawFoundryAgent",
     "RawFoundryAgentChatClient",
     "RawFoundryChatClient",