Execute imported runtime package agent (#1782)

* Bind imported runtime package agent identity (#1781)

* Test imported runtime package selection

Author: chubes4

.github/workflows/agent-task-contracts.yml

@@ -55,8 +55,6 @@ jobs:
       - run: npm run build
       - run: npm run test:agent-task-contracts
       - run: npm run test:runtime-sources-playground-integration
-        env:
-          WP_CODEBOX_RUN_NETWORK_INTEGRATION: "1"
       - run: npm run test:redaction
       - run: npm run test:production-boundary-enforcement
       - run: npm run test:runtime-tool-policy

packages/cli/src/agent-code.ts

@@ -110,11 +110,8 @@ $sandbox_external_runtime_package_import = wp_codebox_import_external_runtime_ag
 $sandbox_stack['external_runtime_package_import'] = $sandbox_external_runtime_package_import;
 if (!empty($sandbox_external_runtime_package_import['success']) && empty($sandbox_external_runtime_package_import['skipped'])) {
     $sandbox_imported_agent = (string) ($sandbox_external_runtime_package_import['identity']['slug'] ?? '');
-    if (is_array($sandbox_runtime_task) && '' !== $sandbox_imported_agent) {
-        $sandbox_runtime_task['input']['package']['slug'] = $sandbox_imported_agent;
-        $sandbox_runtime_task['input']['package']['bootstrap_imported'] = true;
-        $sandbox_runtime_task['input']['metadata']['imported_agent'] = array('slug' => $sandbox_imported_agent);
-    }
+    $sandbox_external_runtime_package_import = wp_codebox_bind_external_runtime_package_identity($sandbox_runtime_task, $sandbox_imported_agent, $sandbox_external_runtime_package_import);
+    $sandbox_stack['external_runtime_package_import'] = $sandbox_external_runtime_package_import;
 }
 
 add_filter('agents_chat_runtime_principal_permission', static function (bool $allowed, $principal, array $input): bool {
@@ -263,6 +260,31 @@ function wp_codebox_import_external_runtime_agent_package(array $runtime_task):
     }
 }
 
+function wp_codebox_bind_external_runtime_package_identity(&$runtime_task, string $agent_slug, array $import): array {
+    if (!is_array($runtime_task) || '' === $agent_slug) {
+        return array('success' => false, 'error' => array('code' => 'wp_codebox_external_runtime_package_identity_missing', 'message' => 'Canonical runtime package import did not produce one agent identity.'));
+    }
+    $task_input = is_array($runtime_task['input'] ?? null) ? $runtime_task['input'] : null;
+    $package = is_array($task_input['package'] ?? null) ? $task_input['package'] : null;
+    $chat_input = is_array($task_input['input'] ?? null) ? $task_input['input'] : array();
+    $metadata = is_array($task_input['metadata'] ?? null) ? $task_input['metadata'] : array();
+    $requested_package_slug = is_array($package) ? (string) ($package['slug'] ?? '') : '';
+    $requested_agent_slug = is_array($chat_input) ? (string) ($chat_input['agent'] ?? '') : '';
+    $metadata_agent_slug = is_array($metadata['imported_agent'] ?? null) ? (string) ($metadata['imported_agent']['slug'] ?? '') : '';
+    if (!is_array($package) || ('' !== $requested_package_slug && !hash_equals($agent_slug, $requested_package_slug)) || ('' !== $requested_agent_slug && !hash_equals($agent_slug, $requested_agent_slug)) || ('' !== $metadata_agent_slug && !hash_equals($agent_slug, $metadata_agent_slug))) {
+        return array('success' => false, 'error' => array('code' => 'wp_codebox_external_runtime_package_identity_mismatch', 'message' => 'Runtime package execution must use the single agent identity verified during canonical import.'));
+    }
+    $package['slug'] = $agent_slug;
+    $package['bootstrap_imported'] = true;
+    $chat_input['agent'] = $agent_slug;
+    $metadata['imported_agent'] = array('slug' => $agent_slug);
+    $task_input['package'] = $package;
+    $task_input['input'] = $chat_input;
+    $task_input['metadata'] = $metadata;
+    $runtime_task['input'] = $task_input;
+    return $import;
+}
+
 function wp_codebox_json_encode_agent_runtime_payload($value): string {
     $json = wp_json_encode($value, JSON_PRETTY_PRINT | JSON_UNESCAPED_SLASHES | JSON_INVALID_UTF8_SUBSTITUTE);
     if (false !== $json) {

packages/wordpress-plugin/src/class-wp-codebox-runtime-package-executor.php

@@ -41,15 +41,17 @@ public function run( array $task ): array|WP_Error {
 			return $imports;
 		}
 
-		$agent_slug = $this->imported_agent_slug( $task );
+		$agent_slug = $this->imported_agent_slug( $task, $imports );
 		if ( is_wp_error( $agent_slug ) ) {
 			return $agent_slug;
 		}
-		if ( '' !== $agent_slug ) {
-			$input           = is_array( $task['input'] ?? null ) ? $task['input'] : array();
-			$input['agent']  = $agent_slug;
-			$task['input']   = $input;
+		$input = is_array( $task['input'] ?? null ) ? $task['input'] : array();
+		$requested_agent = $this->string_value( $input['agent'] ?? '' );
+		if ( '' !== $requested_agent && ! hash_equals( $agent_slug, $requested_agent ) ) {
+			return new WP_Error( 'wp_codebox_runtime_package_agent_identity_mismatch', 'Runtime package execution must use the agent identity verified during import.', array( 'status' => 400 ) );
 		}
+		$input['agent'] = $agent_slug;
+		$task['input']  = $input;
 
 		$result = $this->execute_workflow( $task );
 		if ( is_wp_error( $result ) ) {
@@ -118,9 +120,6 @@ private function execute_workflow( array $task ): array|WP_Error {
 		$ability = 'agents/chat';
 		$input   = is_array( $task['input'] ?? null ) ? $task['input'] : array();
 		$package = is_array( $task['package'] ?? null ) ? $task['package'] : array();
-		if ( ! isset( $input['agent'] ) && '' !== $this->string_value( $package['slug'] ?? '' ) ) {
-			$input['agent'] = $this->string_value( $package['slug'] );
-		}
 		if ( ! isset( $input['message'] ) ) {
 			$input['message'] = $this->string_value( $input['prompt'] ?? '' );
 		}
@@ -149,16 +148,22 @@ private function execute_workflow( array $task ): array|WP_Error {
 		return $result;
 	}
 
-	/** @param array<string,mixed> $task Runtime package task. @return string|WP_Error */
-	private function imported_agent_slug( array $task ): string|WP_Error {
+	/** @param array<string,mixed> $task Runtime package task. @param array<int,array<string,mixed>> $imports Import results. @return string|WP_Error */
+	private function imported_agent_slug( array $task, array $imports ): string|WP_Error {
 		$package = is_array( $task['package'] ?? null ) ? $task['package'] : array();
+		$imported_slugs = array_values( array_unique( array_filter( array_map( static fn( mixed $import ): string => is_array( $import ) && ! empty( $import['success'] ) ? trim( (string) ( $import['agent_slug'] ?? '' ) ) : '', $imports ) ) ) );
+		if ( 1 !== count( $imports ) || 1 !== count( $imported_slugs ) ) {
+			return new WP_Error( 'wp_codebox_runtime_package_imported_agent_unresolved', 'Runtime package import must resolve exactly one agent identity.', array( 'status' => 400 ) );
+		}
+		$slug = $imported_slugs[0];
+		if ( ! hash_equals( $slug, $this->string_value( $package['slug'] ?? '' ) ) || ! function_exists( 'wp_get_agent' ) || ! wp_get_agent( $slug ) ) {
+			return new WP_Error( 'wp_codebox_runtime_package_imported_agent_unresolved', 'The imported runtime package agent identity did not resolve exactly.', array( 'status' => 400 ) );
+		}
 		if ( empty( $package['bootstrap_imported'] ) ) {
-			return '';
+			return $slug;
 		}
-		$metadata = is_array( $task['metadata'] ?? null ) ? $task['metadata'] : array();
-		$slug     = $this->string_value( $metadata['imported_agent']['slug'] ?? '' );
 		$bootstrap = is_array( $GLOBALS['wp_codebox_private_runtime_package_import'] ?? null ) ? $GLOBALS['wp_codebox_private_runtime_package_import'] : array();
-		if ( '' === $slug || ! hash_equals( $slug, $this->string_value( $bootstrap['identity']['slug'] ?? '' ) ) || ! function_exists( 'wp_get_agent' ) || ! wp_get_agent( $slug ) ) {
+		if ( ! hash_equals( $slug, $this->string_value( $bootstrap['identity']['slug'] ?? '' ) ) ) {
 			return new WP_Error( 'wp_codebox_runtime_package_imported_agent_unresolved', 'The imported runtime package agent identity did not resolve exactly.', array( 'status' => 400 ) );
 		}
 

scripts/php-runtime-package-public-contract-smoke.php

@@ -164,13 +164,18 @@ function wp_get_ability( string $name ): ?WP_Codebox_Runtime_Package_Smoke_Abili
 );
 $GLOBALS['wp_codebox_runtime_package_registered_agents']['store-idea-agent'] = true;
 $bootstrap_task = $wpsg_like_task;
-$bootstrap_task['package']['slug']               = 'caller-controlled-agent';
 $bootstrap_task['package']['bootstrap_imported'] = true;
-$bootstrap_task['input']['agent']                = 'caller-controlled-agent';
+$bootstrap_task['input']['agent']                = 'store-idea-agent';
 $bootstrap = WP_Codebox_Abilities::run_runtime_package( $bootstrap_task + array( 'runtime_provider' => 'codebox-runtime-package' ) );
 assert( ! is_wp_error( $bootstrap ) );
 assert( 'store-idea-agent' === $GLOBALS['wp_codebox_runtime_package_smoke_input']['agent'] );
 
+$spoofed_bootstrap_task = $bootstrap_task;
+$spoofed_bootstrap_task['input']['agent'] = 'caller-controlled-agent';
+$spoofed_bootstrap = WP_Codebox_Abilities::run_runtime_package( $spoofed_bootstrap_task + array( 'runtime_provider' => 'codebox-runtime-package' ) );
+assert( is_wp_error( $spoofed_bootstrap ) );
+assert( 'wp_codebox_runtime_package_agent_identity_mismatch' === $spoofed_bootstrap->get_error_code() );
+
 file_put_contents( $staged_bundle_file, "tampered\n" );
 $tampered = WP_Codebox_Abilities::run_runtime_package( $wpsg_like_task + array( 'runtime_provider' => 'codebox-runtime-package' ) );
 assert( is_wp_error( $tampered ) );

tests/agent-no-data-machine-loop.test.ts

@@ -175,16 +175,19 @@ assert.match(sandboxAgentCode, /'model' => \$configured_model/)
 const docsAgentDirectory = process.env.DOCS_AGENT_DIR
 const publicPackageBytes = docsAgentDirectory
   ? await readFile(join(docsAgentDirectory, "bundles", "technical-docs-agent", "native", "technical-docs-maintenance-agent.agent.json"))
-  : await readFile(new URL("./fixtures/external-native-package/flat-agent.agent.json", import.meta.url))
+  : await readFile(new URL("./fixtures/external-native-package/technical-bootstrap-agent.agent.json", import.meta.url))
 const publicAgentSlug = canonicalExternalNativeAgentIdentity(publicPackageBytes).slug
+const skillsPackageBytes = await readFile(new URL("./fixtures/external-native-package/skills-agent.agent.json", import.meta.url))
+assert.deepEqual(canonicalExternalNativeAgentIdentity(skillsPackageBytes), { slug: "skills-agent" })
+assert.deepEqual(canonicalExternalNativeAgentIdentity(publicPackageBytes), { slug: docsAgentDirectory ? "technical-docs-maintenance-agent" : "technical-bootstrap-agent" })
 const publicPackageDigest = `sha256-bytes-v1:${await import("node:crypto").then(({ createHash }) => createHash("sha256").update(publicPackageBytes).digest("hex"))}`
 const bootstrapCode = await resolveSandboxTaskCode({
   task: "Say hello",
   agent: "wp-codebox-sandbox",
   runtimeTask: {
     ability: "wp-codebox/run-runtime-package",
     input: {
-      package: { slug: "caller-controlled-agent", source: "public-external-package", bootstrap: { encoding: "base64", bytes: publicPackageBytes.toString("base64"), digest: publicPackageDigest } },
+      package: { slug: publicAgentSlug, source: "public-external-package", bootstrap: { encoding: "base64", bytes: publicPackageBytes.toString("base64"), digest: publicPackageDigest } },
     },
   },
   sandboxToolPolicy: { schema: "wp-codebox/sandbox-tool-policy/v1", version: 1, tools: [] },
@@ -211,6 +214,30 @@ assert.match(bootstrapCode, /wp_codebox_import_external_runtime_agent_package/)
 assert.ok(bootstrapCode.indexOf("wp_codebox_import_external_runtime_agent_package") < bootstrapCode.indexOf("wp_codebox_resolve_runtime_task_ability"))
 assert.match(bootstrapCode, /base64_decode\(\$bootstrap\['bytes'\], true\)/)
 
+const spoofedBootstrapCode = await resolveSandboxTaskCode({
+  task: "Say hello",
+  agent: "wp-codebox-sandbox",
+  runtimeTask: {
+    ability: "wp-codebox/run-runtime-package",
+    input: {
+      package: { slug: publicAgentSlug, source: "public-external-package", bootstrap: { encoding: "base64", bytes: publicPackageBytes.toString("base64"), digest: publicPackageDigest } },
+      input: { agent: "caller-controlled-agent" },
+    },
+  },
+  sandboxToolPolicy: { schema: "wp-codebox/sandbox-tool-policy/v1", version: 1, tools: [] },
+})
+const spoofedBootstrapOutput = execFileSync("php", ["-r", `${phpPreamble}
+function wp_agent_import_runtime_bundles($bundles, $options) {
+    $package = json_decode((string) file_get_contents((string) ($bundles[0]['source'] ?? '')), true);
+    $slug = $package['agent']['agent_slug'] ?? '';
+    WP_Agents_Registry::get_instance()->register($slug, array('source' => 'canonical-importer'));
+    return array(array('success' => true, 'agent_slug' => $slug));
+}
+${spoofedBootstrapCode}`], { encoding: "utf8" })
+const spoofedBootstrap = JSON.parse(spoofedBootstrapOutput) as { agent_runtime?: { success?: boolean, error?: { code?: string } } }
+assert.equal(spoofedBootstrap.agent_runtime?.success, false)
+assert.equal(spoofedBootstrap.agent_runtime?.error?.code, "wp_codebox_external_runtime_package_identity_mismatch")
+
 const failedImportOutput = execFileSync("php", ["-r", `${phpPreamble}
 function wp_agent_import_runtime_bundles($bundles, $options) { $GLOBALS['external_source'] = $bundles[0]['source']; return array(array('success' => false)); }
 ${bootstrapCode}`], {

tests/fixtures/external-native-package/skills-agent.agent.json

@@ -0,0 +1,15 @@
+{
+  "schema_version": 1,
+  "bundle_slug": "skills-agent",
+  "bundle_version": "1.0.0",
+  "agent": {
+    "agent_slug": "skills-agent",
+    "agent_name": "Skills Agent",
+    "description": "Fixture for a package that supplies skills through enabled tools.",
+    "agent_config": {
+      "instructions": "SKILLS PACKAGE INSTRUCTION",
+      "enabled_tools": ["workspace_read", "workspace_write"],
+      "modes": ["chat", "task"]
+    }
+  }
+}

tests/fixtures/external-native-package/technical-bootstrap-agent.agent.json

@@ -0,0 +1,16 @@
+{
+  "schema_version": 1,
+  "bundle_slug": "technical-bootstrap-agent",
+  "bundle_version": "1.0.0",
+  "agent": {
+    "agent_slug": "technical-bootstrap-agent",
+    "agent_name": "Technical Bootstrap Agent",
+    "description": "Fixture for a technical documentation bootstrap package.",
+    "agent_config": {
+      "instructions": "TECHNICAL BOOTSTRAP PACKAGE INSTRUCTION",
+      "enabled_tools": ["workspace_ls", "workspace_read", "workspace_write"],
+      "tool_call_rules": [{"id": "require-write", "require_tool_use": true, "require_one_of": ["workspace_write"]}],
+      "modes": ["chat"]
+    }
+  }
+}