RSM-3143: research — DLA integration via pi-coding-agent bridge

.buildkite/pipeline.yml

@@ -3,8 +3,16 @@
 
 # IMAGE_ID is an env var that only macOS agents need.
 # Defining it at the root level propagates it too all agents, which can seem unnecessary but is a the same time convenient and DRY.
+#
+# PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 prevents the `playwright install chromium`
+# postinstall (pulled in transitively by the `data-liberation` agent dep) from
+# downloading ~150 MB of Chromium during every CI `npm install` / `npm ci`. The
+# DLA runtime bootstraps Chromium lazily on first use of a Wix/Squarespace
+# adapter, so CI builds never need it. End-user `npm install -g wp-studio` runs
+# the postinstall normally and pays the cost once on first install.
 env:
   IMAGE_ID: $IMAGE_ID
+  PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD: "1"
 
 e2e_config: &e2e_config
   command: bash .buildkite/commands/run-e2e-tests.sh "{{matrix.platform}}" "{{matrix.arch}}"

.buildkite/release-build-and-distribute.yml

@@ -9,9 +9,16 @@
 # Each step checks out the release branch to ensure it builds the latest commit.
 ---
 
-# Used by mac agents only
+# IMAGE_ID is used by mac agents only.
+#
+# PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 prevents the `playwright install chromium`
+# postinstall (pulled in transitively by the `data-liberation` agent dep) from
+# downloading ~150 MB of Chromium during every release-build `npm install`. The
+# DLA runtime bootstraps Chromium lazily on first use of a Wix/Squarespace
+# adapter, so release builds never need it bundled.
 env:
   IMAGE_ID: $IMAGE_ID
+  PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD: "1"
 
 steps:
   - group: 📦 Build for Mac

.buildkite/release-pipelines/code-freeze.yml

@@ -1,8 +1,13 @@
 # yaml-language-server: $schema=https://raw.githubusercontent.com/buildkite/pipeline-schema/main/schema.json
 ---
 
+# PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 prevents the `playwright install chromium`
+# postinstall (pulled in transitively by the `data-liberation` agent dep) from
+# downloading ~150 MB of Chromium during the code-freeze `npm install`. The
+# DLA runtime bootstraps Chromium lazily on first use, so CI never needs it.
 env:
   IMAGE_ID: $IMAGE_ID
+  PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD: "1"
 
 steps:
   - label: ":snowflake: Code Freeze"

.github/workflows/publish-npm-package.yml

@@ -12,6 +12,15 @@ permissions:
   contents: read
   id-token: write # Required for npm trusted publishing (OIDC)
 
+# PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD=1 prevents the `playwright install chromium`
+# postinstall (pulled in transitively by the `data-liberation` agent dep) from
+# downloading ~150 MB of Chromium during `npm ci`. The DLA runtime bootstraps
+# Chromium lazily on first use of a Wix/Squarespace adapter, so the publish job
+# doesn't need it. End-user `npm install -g wp-studio` runs the postinstall
+# normally and pays the cost once on first install.
+env:
+  PLAYWRIGHT_SKIP_BROWSER_DOWNLOAD: "1"
+
 jobs:
   publish:
     runs-on: ubuntu-latest

apps/cli/README.md

@@ -27,6 +27,7 @@ The Studio CLI lets you:
 - [Quick start](#quick-start)
 - [Usage](#usage)
 - [Studio Code](#studio-code)
+- [Migrate from a closed platform](#migrate-from-a-closed-platform)
 - [Import and export](#import-and-export)
 - [Sync with WordPress.com and Pressable](#sync-with-wordpresscom-and-pressable)
 - [Preview sites](#preview-sites)
@@ -98,6 +99,28 @@ studio code sessions list
 studio code sessions resume
 ```
 
+## Migrate from a closed platform
+
+The Studio CLI can move a site off a closed web platform — GoDaddy Websites & Marketing, Hostinger, HubSpot, Shopify, Squarespace, Webflow, Weebly, and Wix — into a fresh local WordPress site. The migration inspects the source, extracts its content into a WXR archive plus media, verifies the result, and lands everything in a new Studio site under `~/Studio/`. Powered by [Data Liberation Agent](https://github.com/Automattic/data-liberation-agent).
+
+The recommended path is the `/liberate` slash command inside `studio code`. The agent walks you through detect, extract, verify, site-create, and import, confirming each heavier step before it runs. This path is gated behind a feature flag while it stabilizes — start `studio code` with `STUDIO_DLA_ENABLED=1` to enable it, then invoke the skill with or without a URL:
+
+```bash
+STUDIO_DLA_ENABLED=1 studio code
+# inside the session:
+# /liberate
+# /liberate https://example.com
+```
+
+For headless or scripted use, run the standalone command instead. It spawns the Data Liberation Agent CLI directly with no agent in the loop, streaming progress straight to your terminal:
+
+```bash
+studio liberate https://example.com
+studio liberate https://example.com --output ./out --non-interactive
+```
+
+Two source platforms need credentials before the extract step: Webflow requires `LIBERATION_TOKEN` (a Webflow site token) and Shopify requires `SHOPIFY_ADMIN_TOKEN` (a Shopify Admin API token with read access to products and orders). Set them in your shell before running either path. Installing the Studio CLI also downloads a Playwright Chromium build (~150 MB) used for the Wix and Squarespace adapters, so the initial `npm install -g wp-studio` pulls more than the base CLI does.
+
 ## Import and export
 
 The Studio CLI allows you to import and export local backups.

apps/cli/ai/runtimes/pi/index.ts

@@ -18,6 +18,7 @@ import {
 	SettingsManager,
 	type AgentSession,
 	type AgentSessionEvent,
+	type ExtensionFactory,
 	type SessionManager,
 	type ToolDefinition,
 } from '@mariozechner/pi-coding-agent';
@@ -27,6 +28,12 @@ import {
 	type AiModelFamily,
 	type AiModelId,
 } from '@studio/common/ai/models';
+import {
+	createDlaPolicyFactory,
+	defaultPolicyBuckets,
+	startDlaBridge,
+	type DlaBridge,
+} from '@studio/dla';
 import { buildSystemPrompt } from 'cli/ai/system-prompt';
 import { resolveStudioToolDefinitions } from 'cli/ai/tools';
 import { createAskUserQuestionTool } from 'cli/ai/tools/ask-user-question';
@@ -203,8 +210,9 @@ async function runAgentSessionTurn(
 
 	let session: AgentSession | undefined;
 	let unsubscribe: ( () => void ) | undefined;
+	const dlaBridge = await maybeStartDlaBridge( config );
 	try {
-		session = await createStudioAgentSession( config, family, resolved.creds );
+		session = await createStudioAgentSession( config, family, resolved.creds, dlaBridge );
 		setActiveSession( session );
 		unsubscribe = session.subscribe( config.onEvent );
 
@@ -223,13 +231,61 @@ async function runAgentSessionTurn(
 		unsubscribe?.();
 		session?.dispose();
 		setActiveSession( undefined );
+		await dlaBridge?.dispose();
+	}
+}
+
+/**
+ * Bring up the Data Liberation Agent (DLA) MCP bridge for this session,
+ * gated on `STUDIO_DLA_ENABLED === '1'`.
+ *
+ * Failures to spawn or connect the bridge are non-fatal: a warning is
+ * logged and `undefined` is returned so the session still starts with the
+ * usual Studio tools. The bridge itself also degrades gracefully (returning
+ * `degraded: true` with an empty tool list) when `listTools` fails — in
+ * that case the bridge handle is still returned so the caller can dispose
+ * it during teardown.
+ *
+ * The wpcom access token, when present in the session config, is forwarded
+ * to the DLA child as `STUDIO_WPCOM_TOKEN` so DLA can authenticate against
+ * WordPress.com APIs when surfacing remote-site flows.
+ *
+ * @param config - The resolved session config for this turn.
+ * @returns A `DlaBridge` handle, or `undefined` when DLA is disabled or
+ *   the bridge could not be constructed at all.
+ */
+async function maybeStartDlaBridge(
+	config: ResolvedStudioAgentTurnConfig
+): Promise< DlaBridge | undefined > {
+	if ( config.env.STUDIO_DLA_ENABLED !== '1' ) {
+		return undefined;
+	}
+	try {
+		const bridge = await startDlaBridge( {
+			wpcomToken: config.wpcomAccessToken,
+		} );
+		if ( bridge.degraded ) {
+			console.warn(
+				`[studio code] DLA bridge degraded; continuing without DLA tools (${
+					bridge.degradationReason ?? 'unknown reason'
+				}).`
+			);
+		}
+		return bridge;
+	} catch ( error ) {
+		const reason = error instanceof Error ? error.message : String( error );
+		console.warn(
+			`[studio code] DLA bridge failed to start; continuing without DLA tools (${ reason }).`
+		);
+		return undefined;
 	}
 }
 
 async function createStudioAgentSession(
 	config: ResolvedStudioAgentTurnConfig,
 	family: AiModelFamily,
-	creds: ResolvedCredentials
+	creds: ResolvedCredentials,
+	dlaBridge?: DlaBridge
 ): Promise< AgentSession > {
 	const model = buildModel( config.model, family, creds );
 	const isRemoteSite = Boolean( config.activeSite?.remote && config.activeSite?.wpcomSiteId );
@@ -249,10 +305,10 @@ async function createStudioAgentSession(
 			: { previewSteering: isForkedByDesktop, remoteSession }
 	);
 
-	const tools = buildAgentTools( config, isForkedByDesktop, remoteSession );
-	const toolDefinitions = tools.map( toToolDefinition );
+	const toolDefinitions = buildAgentTools( config, isForkedByDesktop, remoteSession, dlaBridge );
 	const { authStorage, modelRegistry } = createModelRegistry( model, family, creds );
 	const settingsManager = createSettingsManager( config.env );
+	const extensionFactories = resolveDlaExtensionFactories( config.env );
 	const resourceLoader = new DefaultResourceLoader( {
 		cwd: STUDIO_SITES_ROOT,
 		agentDir: STUDIO_AGENT_DIR,
@@ -263,6 +319,7 @@ async function createStudioAgentSession(
 		noThemes: true,
 		noContextFiles: true,
 		systemPrompt,
+		extensionFactories,
 	} );
 	await resourceLoader.reload();
 
@@ -387,6 +444,33 @@ function createSettingsManager( _env: Record< string, string > ): SettingsManage
 	} );
 }
 
+/**
+ * Build the list of pi `ExtensionFactory` values to feed into
+ * `DefaultResourceLoader.extensionFactories` based on the session's env.
+ *
+ * Inline `extensionFactories` are loaded even when `noExtensions: true`
+ * is set on `DefaultResourceLoader` (verified at `resource-loader.js`'s
+ * `loadExtensionFactories` path), so the runtime keeps the rest of the
+ * extension discovery surface disabled while still mounting the DLA
+ * policy hook when the feature flag is on.
+ *
+ * The DLA policy factory is only mounted when
+ * `STUDIO_DLA_ENABLED === '1'`. v1 ships with the flag off by default;
+ * the same flag also gates the DLA bridge spawn in `runAgentSessionTurn`
+ * (handled separately) so the policy hook is a no-op when the bridge is
+ * inactive.
+ *
+ * @param env - The resolved process env for this session.
+ * @returns An array of factories to forward to `DefaultResourceLoader`.
+ *   Empty when DLA is disabled.
+ */
+function resolveDlaExtensionFactories( env: Record< string, string > ): ExtensionFactory[] {
+	if ( env.STUDIO_DLA_ENABLED !== '1' ) {
+		return [];
+	}
+	return [ createDlaPolicyFactory( defaultPolicyBuckets ) ];
+}
+
 function toToolDefinition( tool: AgentToolAny ): ToolDefinition {
 	return {
 		name: tool.name,
@@ -400,11 +484,34 @@ function toToolDefinition( tool: AgentToolAny ): ToolDefinition {
 	};
 }
 
+/**
+ * Build the `customTools` list passed to pi's `createAgentSession`. The
+ * Studio-native tools (returned as `AgentTool` values) are converted to
+ * pi's `ToolDefinition` shape via `toToolDefinition`, and the bridged DLA
+ * tools — already `ToolDefinition[]` — are spliced into the local-site
+ * tool list when `dlaBridge` is supplied.
+ *
+ * The DLA tools intentionally do not appear in the remote-site branch
+ * because the remote flow already steers callers toward WordPress.com
+ * APIs (`wpcom_request`), and bridging DLA there risks recursive migration
+ * loops back into Studio itself.
+ *
+ * @param config - The resolved per-turn session config.
+ * @param enablePreviewSteering - Whether the runtime is forked by the
+ *   desktop app (which steers users toward the Preview Site flow).
+ * @param remoteSession - Whether this is a remote, WPCOM-backed session.
+ * @param dlaBridge - Optional DLA bridge handle; when present, its
+ *   `tools` are spliced into the local-site tool list. When absent
+ *   (e.g. `STUDIO_DLA_ENABLED` is unset or the bridge failed to spawn)
+ *   the result matches the legacy tool list exactly.
+ * @returns The fully assembled `ToolDefinition[]` for the session.
+ */
 function buildAgentTools(
 	config: ResolvedStudioAgentTurnConfig,
 	enablePreviewSteering: boolean,
-	remoteSession: boolean
-): AgentToolAny[] {
+	remoteSession: boolean,
+	dlaBridge?: DlaBridge
+): ToolDefinition[] {
 	const isRemoteSite = Boolean(
 		config.activeSite?.remote && config.activeSite?.wpcomSiteId && config.wpcomAccessToken
 	);
@@ -417,14 +524,15 @@ function buildAgentTools(
 	const skillTool: AgentToolAny[] = skillToolDef ? [ skillToolDef ] : [];
 
 	if ( isRemoteSite ) {
-		return [
+		const remoteTools: AgentToolAny[] = [
 			createWpcomRequestTool( config.wpcomAccessToken!, config.activeSite!.wpcomSiteId! ),
 			takeScreenshotTool,
 			createSiteTool,
 			pullSiteTool,
 			...askUserTool,
 			...skillTool,
 		];
+		return remoteTools.map( toToolDefinition );
 	}
 
 	const renameTool = ( tool: AgentToolAny, name: string ): AgentToolAny => ( {
@@ -442,12 +550,13 @@ function buildAgentTools(
 		renameTool( createFindTool( STUDIO_SITES_ROOT ), 'Glob' ),
 		renameTool( createLsTool( STUDIO_SITES_ROOT ), 'Ls' ),
 	];
-	return [
+	const studioTools: AgentToolAny[] = [
 		...resolveStudioToolDefinitions( { enablePreviewSteering, remoteSession } ),
 		...askUserTool,
 		...skillTool,
 		...piTools,
 	];
+	return [ ...studioTools.map( toToolDefinition ), ...( dlaBridge?.tools ?? [] ) ];
 }
 
 function parseJsonHeaderEnv( value: string | undefined ): Record< string, string > | undefined {