feat: transfer Foundry provisioning to projects (#9133)

Author: huimiu

.github/workflows/test-ext-azure-ai-agents.yml

@@ -5,6 +5,7 @@ on:
     paths:
       - "cli/azd/extensions/azure.ai.agents/**"
       - "cli/azd/extensions/azure.ai.inspector/**"
+      - "cli/azd/extensions/azure.ai.projects/**"
       - "cli/azd/test/functional/ai_agents/**"
       - ".github/workflows/test-ext-azure-ai-agents.yml"
     branches: [main]
@@ -121,6 +122,20 @@ jobs:
           azd x publish
           azd extension install azure.ai.inspector --source local --force --no-prompt
 
+      - name: Build & install projects dependency from source
+        working-directory: cli/azd/extensions/azure.ai.projects
+        run: |
+          pwsh -File ci-build.ps1 -OutputFileName azure-ai-projects-linux-amd64 -Version "$(cat version.txt | tr -d '\r\n')"
+          mkdir -p bin
+          mv azure-ai-projects-linux-amd64 bin/
+          azd x pack
+          azd x publish
+          azd extension install azure.ai.projects --source local --force --no-prompt
+
+      - name: Run projects unit tests
+        working-directory: cli/azd/extensions/azure.ai.projects
+        run: go test ./... -count=1
+
       - name: Build extension (ci-build.ps1)
         working-directory: cli/azd/extensions/azure.ai.agents
         run: |
@@ -137,7 +152,7 @@ jobs:
         run: |
           azd x pack
           azd x publish
-          azd extension install azure.ai.agents --source local --force --no-prompt
+          azd extension install azure.ai.agents --source local --force --no-dependencies --no-prompt
 
       - name: Run Tier 0 tests (offline, production extension)
         working-directory: cli/azd

cli/azd/extensions/azure.ai.agents/AGENTS.md

@@ -124,6 +124,10 @@ Define new codes in `internal/exterrors/codes.go`.
 
 A new extension release ships in two PRs:
 
+### Provider handoff release
+
+The release that removes `microsoft.foundry` from this extension must be coordinated with the provider-bearing `azure.ai.projects` release. Publish both artifacts before updating either registry entry, then update both entries and the `microsoft.foundry` meta-package together.
+
 ### PR 1 — Version bump
 
 Bumps the extension to the new version. Touches only:

cli/azd/extensions/azure.ai.agents/README.md

@@ -80,11 +80,7 @@ Details:
 
 ## Private networking for `host: azure.ai.project`
 
-Foundry project services can be provisioned as network-secured, VNet-bound
-accounts by adding a `network:` block to the `host: azure.ai.project` service in
-`azure.yaml`. See [Private networking for `host: azure.ai.project`](docs/private-networking.md)
-for the schema reference, BYO-image requirements, and VNet deployment
-cheatsheet.
+Foundry project services can be provisioned as network-secured, VNet-bound accounts by adding a `network:` block to the `host: azure.ai.project` service in `azure.yaml`. The `azure.ai.projects` extension owns that service and the `microsoft.foundry` provider; this extension still authors the block during agent init. See [Private networking for `host: azure.ai.project`](docs/private-networking.md) for the schema reference, BYO-image requirements, and VNet deployment cheatsheet.
 
 ## Local Development
 

cli/azd/extensions/azure.ai.agents/docs/private-networking.md

@@ -4,6 +4,8 @@ A Foundry project service can be provisioned as a **network-secured (VNet-bound)
 
 Do **not** place `network:` on `host: azure.ai.agent`. Agent services describe deployable agents and depend on the project through `uses:`; the project service owns account-level provisioning inputs such as `endpoint:`, `deployments:`, and `network:`.
 
+The `azure.ai.projects` extension owns the project service and the `microsoft.foundry` provider. `azd ai agent init` continues to author the block and eject its IaC during the staged ownership migration.
+
 When `network:` is present, azd always provisions an **account private endpoint** and disables public data-plane access. Dependent stores (Cosmos DB, AI Search, Storage) stay platform-managed.
 
 ```yaml

cli/azd/extensions/azure.ai.agents/extension.yaml

@@ -10,22 +10,19 @@ requiredAzdVersion: ">=1.27.1"
 dependencies:
   - id: azure.ai.inspector
     version: "~1.0.0-beta.1"
+  - id: azure.ai.projects
+    version: "~1.0.0-beta.3"
 language: go
 capabilities:
   - custom-commands
   - lifecycle-events
   - mcp-server
   - service-target-provider
-  - provisioning-provider
-  - validation-provider
   - metadata
 providers:
   - name: azure.ai.agent
     type: service-target
     description: Deploys agents to the Foundry Agent Service
-  - name: microsoft.foundry
-    type: provisioning-provider
-    description: Provisions a Microsoft Foundry project from azure.yaml without an on-disk infra/ directory
 examples:
   - name: init
     description: Initialize a new AI agent project.

cli/azd/extensions/azure.ai.agents/internal/cmd/init_infra.go

@@ -614,9 +614,27 @@ func writeTfvarsFile(infraDir string, params map[string]any) (ejectArtifact, err
 	if v, ok := params["deployments"]; ok {
 		doc["deployments"] = v
 	}
-	if v, ok := params["connections"]; ok {
-		doc["connections"] = v
+	connections, ok := params["connections"].([]synthesis.Connection)
+	if !ok {
+		return ejectArtifact{}, exterrors.Internal(
+			exterrors.CodeInfraEjectWriteFailed,
+			fmt.Sprintf("connections parameter has unexpected type %T", params["connections"]),
+		)
+	}
+	connectionCredentials, ok := params["connectionCredentials"].(map[string]map[string]any)
+	if !ok {
+		return ejectArtifact{}, exterrors.Internal(
+			exterrors.CodeInfraEjectWriteFailed,
+			fmt.Sprintf(
+				"connectionCredentials parameter has unexpected type %T",
+				params["connectionCredentials"],
+			),
+		)
 	}
+	doc["connections"] = synthesis.JoinConnectionCredentials(
+		connections,
+		connectionCredentials,
+	)
 
 	data, err := json.MarshalIndent(doc, "", "  ")
 	if err != nil {

cli/azd/extensions/azure.ai.agents/internal/cmd/init_infra_test.go

@@ -318,9 +318,8 @@ services:
 
 func TestEjectInfra_EjectsConnectionServices(t *testing.T) {
 	// See TestEjectInfra_HappyPath_WritesExpectedFiles for why this is not parallel.
-	// A host: azure.ai.connection service must be synthesized into the
-	// connections param, its module copied into the ejected tree, and any
-	// ${VAR} in credentials kept verbatim (environment-portable).
+	// Connection metadata and credentials are ejected separately.
+	// Bicep keeps credential values in a secure object parameter.
 	dir := t.TempDir()
 	mustWriteFile(t, filepath.Join(dir, "azure.yaml"), `name: my-project
 services:
@@ -374,17 +373,20 @@ services:
 	assert.Equal(t, "CognitiveSearch", conn["category"])
 	assert.Equal(t, "ApiKey", conn["authType"])
 
-	// ${VAR} in credentials must be preserved verbatim on the eject path.
-	creds, ok := conn["credentials"].(map[string]any)
-	require.True(t, ok, "credentials should be an object, got %T", conn["credentials"])
-	assert.Equal(t, "${SEARCH_API_KEY}", creds["key"])
+	assert.NotContains(t, conn, "credentials")
 
 	// Nested CustomKeys credentials must remain an object so Terraform's
 	// optional(any) value can preserve mixed connection credential shapes.
 	mcpConn, ok := conns[0].(map[string]any)
 	require.True(t, ok, "connection entry should be an object, got %T", conns[0])
 	assert.Equal(t, "mcp-conn", mcpConn["name"])
-	mcpCreds := mcpConn["credentials"].(map[string]any)
+	assert.NotContains(t, mcpConn, "credentials")
+
+	secureCreds, ok := doc.Parameters["connectionCredentials"].Value.(map[string]any)
+	require.True(t, ok, "connectionCredentials should be an object")
+	searchCreds := secureCreds["search-conn"].(map[string]any)
+	assert.Equal(t, "${SEARCH_API_KEY}", searchCreds["key"])
+	mcpCreds := secureCreds["mcp-conn"].(map[string]any)
 	keys := mcpCreds["keys"].(map[string]any)
 	assert.Equal(t, "${MCP_KEY}", keys["x-api-key"])
 }
@@ -742,6 +744,7 @@ func TestEjectInfra_Terraform_TfvarsShape(t *testing.T) {
 	conns, ok := doc["connections"].([]any)
 	require.True(t, ok, "connections should be an array, got %T", doc["connections"])
 	assert.Empty(t, conns)
+	assert.NotContains(t, doc, "connectionCredentials")
 }
 
 func TestEjectInfra_Terraform_EjectsConnectionServices(t *testing.T) {
@@ -793,6 +796,7 @@ services:
 	creds, ok := conn["credentials"].(map[string]any)
 	require.True(t, ok, "credentials should be an object, got %T", conn["credentials"])
 	assert.Equal(t, "${SEARCH_API_KEY}", creds["key"])
+	assert.NotContains(t, doc, "connectionCredentials")
 
 	// outputs.tf always carries the connection-names output, unconditional on
 	// includeAcr (unlike the ACR outputs).

cli/azd/extensions/azure.ai.agents/internal/cmd/listen.go

@@ -36,20 +36,6 @@ func configureExtensionHost(host *azdext.ExtensionHost) {
 		WithServiceTarget(AiAgentHost, func() azdext.ServiceTargetProvider {
 			return project.NewAgentServiceTargetProvider(azdClient)
 		}).
-		WithProvisioningProvider(project.FoundryProviderName, func() azdext.ProvisioningProvider {
-			return project.NewFoundryProvisioningProvider(azdClient)
-		}).
-		WithValidationCheck(azdext.ValidationCheckRegistration{
-			// Provider-agnostic check: runs before provisioning regardless of the
-			// provisioning provider. This matters because the azure.ai.agents
-			// extension provisions through its own microsoft.foundry provider,
-			// which never triggers the Bicep-only "arm-provision" checks.
-			CheckType: azdext.ValidationCheckTypeProvision,
-			RuleID:    project.ResourceGroupLocationRuleID,
-			Factory: func() azdext.ValidationCheckProvider {
-				return project.NewResourceGroupLocationCheck(azdClient)
-			},
-		}).
 		WithProjectEventHandler("preprovision", func(ctx context.Context, args *azdext.ProjectEventArgs) error {
 			return preprovisionHandler(ctx, azdClient, args)
 		}).
@@ -68,8 +54,11 @@ func configureExtensionHost(host *azdext.ExtensionHost) {
 }
 
 func preprovisionHandler(ctx context.Context, azdClient *azdext.AzdClient, args *azdext.ProjectEventArgs) error {
-	deployments, err := collectProjectDeployments(args.Project.Services)
-	if err != nil {
+	if err := updateLegacyProjectDeployments(
+		ctx,
+		azdClient,
+		args.Project.Services,
+	); err != nil {
 		return err
 	}
 	connections, err := collectConnections(args.Project.Services)
@@ -83,7 +72,13 @@ func preprovisionHandler(ctx context.Context, azdClient *azdext.AzdClient, args
 			if err := populateContainerSettings(ctx, azdClient, svc); err != nil {
 				return fmt.Errorf("failed to populate container settings for service %q: %w", svc.Name, err)
 			}
-			if err := envUpdate(ctx, azdClient, args.Project, svc, deployments, connections); err != nil {
+			if err := envUpdate(
+				ctx,
+				azdClient,
+				args.Project,
+				svc,
+				connections,
+			); err != nil {
 				return fmt.Errorf("failed to update environment for service %q: %w", svc.Name, err)
 			}
 		}
@@ -168,6 +163,34 @@ func currentEnvName(ctx context.Context, azdClient *azdext.AzdClient) (string, e
 	return resp.Environment.Name, nil
 }
 
+func updateLegacyProjectDeployments(
+	ctx context.Context,
+	azdClient *azdext.AzdClient,
+	services map[string]*azdext.ServiceConfig,
+) error {
+	deployments, err := collectLegacyProjectDeployments(services)
+	if err != nil {
+		return err
+	}
+	if len(deployments) == 0 {
+		return nil
+	}
+
+	envName, err := currentEnvName(ctx, azdClient)
+	if err != nil {
+		return fmt.Errorf(
+			"resolving environment for legacy deployments: %w",
+			err,
+		)
+	}
+	return deploymentEnvUpdate(
+		ctx,
+		deployments,
+		azdClient,
+		envName,
+	)
+}
+
 // developerRBACOnce ensures CheckDeveloperRBAC runs at most once per extension
 // process lifetime. Service-level predeploy handlers fire per-service, but the
 // RBAC pre-flight check is project-scoped and idempotent — running it once is
@@ -194,8 +217,11 @@ func predeployHandler(ctx context.Context, azdClient *azdext.AzdClient, args *az
 		warnDuplicateAgentNames(args.Project)
 	})
 
-	deployments, err := collectProjectDeployments(args.Project.Services)
-	if err != nil {
+	if err := updateLegacyProjectDeployments(
+		ctx,
+		azdClient,
+		args.Project.Services,
+	); err != nil {
 		return err
 	}
 	connections, err := collectConnections(args.Project.Services)
@@ -206,7 +232,13 @@ func predeployHandler(ctx context.Context, azdClient *azdext.AzdClient, args *az
 	if err := populateContainerSettings(ctx, azdClient, svc); err != nil {
 		return fmt.Errorf("failed to populate container settings for service %q: %w", svc.Name, err)
 	}
-	if err := envUpdate(ctx, azdClient, args.Project, svc, deployments, connections); err != nil {
+	if err := envUpdate(
+		ctx,
+		azdClient,
+		args.Project,
+		svc,
+		connections,
+	); err != nil {
 		return fmt.Errorf("failed to update environment for service %q: %w", svc.Name, err)
 	}
 
@@ -477,7 +509,6 @@ func envUpdate(
 	azdClient *azdext.AzdClient,
 	azdProject *azdext.ProjectConfig,
 	svc *azdext.ServiceConfig,
-	deployments []project.Deployment,
 	connections []project.Connection,
 ) error {
 
@@ -495,15 +526,6 @@ func envUpdate(
 		return err
 	}
 
-	// Deployments and connections are sourced from the sibling
-	// azure.ai.project and azure.ai.connection services. Resources and tool
-	// connections stay on the agent service.
-	if len(deployments) > 0 {
-		if err := deploymentEnvUpdate(ctx, deployments, azdClient, currentEnvResponse.Environment.Name); err != nil {
-			return err
-		}
-	}
-
 	if foundryAgentConfig != nil && len(foundryAgentConfig.Resources) > 0 {
 		if err := resourcesEnvUpdate(ctx, foundryAgentConfig.Resources, azdClient, currentEnvResponse.Environment.Name); err != nil {
 			return err

cli/azd/extensions/azure.ai.agents/internal/cmd/resource_services.go

@@ -346,36 +346,23 @@ func reserveServiceName(used map[string]string, name, source string) error {
 	return nil
 }
 
-// collectProjectDeployments gathers the model deployments declared across all
-// azure.ai.project services so provisioning handlers can source them from the
-// sibling project service instead of the agent service config. Services are
-// visited in sorted name order so serialized env-var output stays stable.
-//
-// Falls back to the deployments bundled on the agent service when no project
-// service carries any, so an azure.yaml written before the per-resource split
-// still provisions without re-running init.
-func collectProjectDeployments(services map[string]*azdext.ServiceConfig) ([]project.Deployment, error) {
-	var out []project.Deployment
-	for _, svc := range sortedServices(services) {
-		props := project.ServiceConfigProps(svc)
-		if svc.Host != AiProjectHost || props == nil {
-			continue
-		}
-		var cfg *project.ServiceTargetAgentConfig
-		if err := project.UnmarshalStruct(props, &cfg); err != nil {
-			return nil, fmt.Errorf("parsing project service %q config: %w", svc.Name, err)
-		}
-		if cfg != nil {
-			out = append(out, cfg.Deployments...)
+// collectLegacyProjectDeployments reads only pre-split agent config.
+// A split project disables this compatibility path because projects
+// owns that service's runtime projection.
+func collectLegacyProjectDeployments(
+	services map[string]*azdext.ServiceConfig,
+) ([]project.Deployment, error) {
+	for _, svc := range services {
+		if svc.GetHost() == AiProjectHost {
+			return nil, nil
 		}
 	}
-	if len(out) > 0 {
-		return out, nil
-	}
+
 	legacy, err := collectLegacyAgentConfigs(services)
 	if err != nil {
 		return nil, err
 	}
+	var out []project.Deployment
 	for _, cfg := range legacy {
 		out = append(out, cfg.Deployments...)
 	}