@playdrop/playdrop-cli 0.11.0 → 0.11.4

package/config/client-meta.json

@@ -1,5 +1,5 @@
 {
-  "version": "0.11.0",
+  "version": "0.11.4",
   "build": 1,
   "runtimeSdkVersion": "0.10.0",
   "clients": {

package/dist/apps/validate.js

@@ -48,7 +48,6 @@ const SDK_REFERENCE_PATTERNS = [
 const SDK_INIT_PATTERNS = [
     /\bwindow\s*\.\s*playdrop\s*\.\s*init\s*\(/g,
     /\bplaydrop\s*\.\s*init\s*\(/g,
-    /\bsdk\s*\.\s*initialize\s*\(/g,
 ];
 const SDK_READY_PATTERNS = [
     /\bsdk\s*\??\.\s*host\s*\??\.\s*ready\s*\(/g,

package/dist/captureRuntime.js

@@ -16,6 +16,7 @@ const FRAME_SELECTOR = 'iframe[title="Game"]';
 exports.CAPTURE_LOG_LEVEL_VALUES = ['debug', 'info', 'warn', 'error'];
 exports.MAX_CAPTURE_TIMEOUT_SECONDS = 600;
 const GOOGLE_TELEMETRY_HOSTS = [
+    'analytics.google.com',
     'googletagmanager.com',
     'google-analytics.com',
     'googleadservices.com',
@@ -97,7 +98,13 @@ function extractRefusedResourceUrls(text) {
 }
 function isKnownGoogleTelemetryConsoleIssue(text, locationUrl) {
     if (/Content Security Policy/i.test(text)) {
-        return extractRefusedResourceUrls(text).some(isKnownGoogleTelemetryUrl);
+        const refusedUrls = extractRefusedResourceUrls(text);
+        if (refusedUrls.some(isKnownGoogleTelemetryUrl)) {
+            return true;
+        }
+        const messageUrls = extractHttpUrls(text);
+        return messageUrls.some(isKnownGoogleTelemetryUrl)
+            && !messageUrls.some(isLoopbackOrLocalUrl);
     }
     if (!/ERR_BLOCKED_BY_CLIENT/i.test(text)) {
         return false;

package/dist/commands/captureRemote.d.ts

@@ -9,6 +9,7 @@ type CaptureRemoteOptions = {
     loginUrl?: string;
     actions?: string;
     viewport?: string;
+    surface?: string;
     fpsSampleMs?: string | number;
 };
 export declare function buildCaptureRemoteCommandArgs(url: string | undefined, options?: CaptureRemoteOptions): string[] | null;

package/dist/commands/captureRemote.js

@@ -10,6 +10,29 @@ const http_1 = require("../http");
 const messages_1 = require("../messages");
 const devShared_1 = require("./devShared");
 const captureRuntime_1 = require("../captureRuntime");
+const MOBILE_USER_AGENT = 'Mozilla/5.0 (iPhone; CPU iPhone OS 16_0 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/16.0 Mobile/15E148 Safari/604.1';
+const SURFACE_CONTEXT_OPTIONS = {
+    desktop: {
+        viewport: { width: 1280, height: 720 },
+        deviceScaleFactor: 1,
+        isMobile: false,
+        hasTouch: false,
+    },
+    'mobile-landscape': {
+        viewport: { width: 896, height: 414 },
+        deviceScaleFactor: 2,
+        isMobile: true,
+        hasTouch: true,
+        userAgent: MOBILE_USER_AGENT,
+    },
+    'mobile-portrait': {
+        viewport: { width: 414, height: 896 },
+        deviceScaleFactor: 2,
+        isMobile: true,
+        hasTouch: true,
+        userAgent: MOBILE_USER_AGENT,
+    },
+};
 function readUnsupportedPlaydropCaptureMessage(targetUrl) {
     let parsed;
     try {
@@ -69,6 +92,16 @@ function parseViewport(raw) {
     }
     return { width, height };
 }
+function parseSurface(raw) {
+    const normalized = raw?.trim().toLowerCase();
+    if (!normalized) {
+        return null;
+    }
+    if (normalized === 'desktop' || normalized === 'mobile-landscape' || normalized === 'mobile-portrait') {
+        return normalized;
+    }
+    throw new Error('invalid_surface');
+}
 function parseFpsSampleMs(raw) {
     if (raw === undefined) {
         return null;
@@ -172,6 +205,7 @@ function buildCaptureRemoteCommandArgs(url, options = {}) {
     pushOptionalArg(args, '--expected-url', options.expectedUrl);
     pushOptionalArg(args, '--actions', options.actions);
     pushOptionalArg(args, '--viewport', options.viewport);
+    pushOptionalArg(args, '--surface', options.surface);
     if (options.fpsSampleMs !== undefined) {
         args.push('--fps-sample-ms', String(parseFpsSampleMs(options.fpsSampleMs)));
     }
@@ -205,6 +239,7 @@ function parseOptions(url, options = {}) {
         logPath: options.log?.trim() ? (0, node_path_1.resolve)(process.cwd(), options.log.trim()) : null,
         actionsPath: options.actions?.trim() ? (0, node_path_1.resolve)(process.cwd(), options.actions.trim()) : null,
         viewport: parseViewport(options.viewport),
+        surface: parseSurface(options.surface),
         fpsSampleMs: parseFpsSampleMs(options.fpsSampleMs),
         login: username
             ? {
@@ -247,6 +282,11 @@ async function captureRemote(url, options = {}) {
             process.exitCode = 1;
             return;
         }
+        if (error instanceof Error && error.message === 'invalid_surface') {
+            (0, messages_1.printErrorWithHelp)('The --surface value must be desktop, mobile-landscape, or mobile-portrait.', [], { command: 'project capture remote' });
+            process.exitCode = 1;
+            return;
+        }
         if (error instanceof Error && error.message === 'invalid_fps_sample_ms') {
             (0, messages_1.printErrorWithHelp)('The --fps-sample-ms value must be an integer between 100 and 60000 milliseconds.', [], { command: 'project capture remote' });
             process.exitCode = 1;
@@ -318,7 +358,11 @@ async function captureRemote(url, options = {}) {
                 user: parsed.login ? undefined : currentUser,
                 login: parsed.login,
                 actions: actions ?? undefined,
-                contextOptions: parsed.viewport ? { viewport: parsed.viewport } : undefined,
+                contextOptions: parsed.surface
+                    ? { ...SURFACE_CONTEXT_OPTIONS[parsed.surface], ...(parsed.viewport ? { viewport: parsed.viewport } : {}) }
+                    : parsed.viewport
+                        ? { viewport: parsed.viewport }
+                        : undefined,
                 frameRateSampleMs: parsed.fpsSampleMs ?? undefined,
             });
             if (result.errorCount > 0) {

package/dist/commands/worker.d.ts

@@ -77,6 +77,7 @@ type TaskCatalogueReportOptions = {
 type TaskCompleteOptions = {
     env?: string;
     nextSteps?: string;
+    summary?: string;
 };
 type TaskClaimSlugOptions = {
     appName?: string;
@@ -98,6 +99,10 @@ type TaskSubmitReviewOptions = {
     evidenceDir?: string;
     nextSteps?: string;
 };
+type TaskSubmitEvalOptions = {
+    env?: string;
+    resultFile?: string;
+};
 export type WorkerHealthAlertInput = {
     state: 'started' | 'stopped' | 'crashed';
     env: string;
@@ -157,6 +162,10 @@ export declare function stagePlaydropPluginReferences(input: {
     pluginRoot: string;
     kind: AgentTaskKind;
 }): Promise<string[]>;
+export declare function stageAssignmentPluginBundle(input: {
+    workspaceDir: string;
+    pluginBundle: NonNullable<WorkerAgentTaskAssignmentResponse['pluginBundle']>;
+}): Promise<string>;
 export declare function discoverWorkerProjectRoot(workspaceDir: string): string;
 export declare function readWorkerTaskState(): WorkerTaskState | null;
 export declare function buildWorkerHealthAlertText(input: WorkerHealthAlertInput): string;
@@ -181,6 +190,7 @@ export declare function buildWorkerCapabilities(input: string | {
     } | null;
     maxParallelTasks?: number | null;
     runningTaskCount?: number | null;
+    providerAccountLabel?: string | null;
 }): AgentWorkerCapabilities;
 export declare function buildWorkerClaimBody(input: {
     workerKey: string;
@@ -254,4 +264,5 @@ export declare function readReviewEvidenceFiles(evidenceDir: string | undefined)
     contentBase64: string;
 }>;
 export declare function submitReviewTask(options: TaskSubmitReviewOptions): Promise<void>;
+export declare function submitEvalTask(options: TaskSubmitEvalOptions): Promise<void>;
 export declare function failTask(options: TaskFailOptions): Promise<void>;

package/dist/commands/worker.js

@@ -18,6 +18,7 @@ exports.assertTaskUploadResultMatchesContext = assertTaskUploadResultMatchesCont
 exports.stageAssignmentWorkspaceFiles = stageAssignmentWorkspaceFiles;
 exports.resolvePlaydropPluginRoot = resolvePlaydropPluginRoot;
 exports.stagePlaydropPluginReferences = stagePlaydropPluginReferences;
+exports.stageAssignmentPluginBundle = stageAssignmentPluginBundle;
 exports.discoverWorkerProjectRoot = discoverWorkerProjectRoot;
 exports.readWorkerTaskState = readWorkerTaskState;
 exports.buildWorkerHealthAlertText = buildWorkerHealthAlertText;
@@ -52,6 +53,7 @@ exports.claimSlugTask = claimSlugTask;
 exports.completeTask = completeTask;
 exports.readReviewEvidenceFiles = readReviewEvidenceFiles;
 exports.submitReviewTask = submitReviewTask;
+exports.submitEvalTask = submitEvalTask;
 exports.failTask = failTask;
 const promises_1 = require("node:fs/promises");
 const node_fs_1 = require("node:fs");
@@ -59,6 +61,7 @@ const node_os_1 = __importDefault(require("node:os"));
 const node_path_1 = __importDefault(require("node:path"));
 const node_process_1 = __importDefault(require("node:process"));
 const node_module_1 = require("node:module");
+const node_crypto_1 = require("node:crypto");
 const fflate_1 = require("fflate");
 const types_1 = require("@playdrop/types");
 const commandContext_1 = require("../commandContext");
@@ -99,10 +102,13 @@ const FAIL_REPORT_RETRY_DELAY_MS = 2000;
 const WORKER_TRANSCRIPT_CHUNK_BATCH_SIZE = 100;
 const SLACK_API_BASE_URL = 'https://slack.com/api';
 const WORKER_WRAPPER_CONTRACT_VERSION = 1;
-const WORKER_SUPPORTED_KINDS = ['NEW_GAME', 'REMIX_GAME', 'GAME_UPDATE', 'GAME_REVIEW'];
+const WORKER_SUPPORTED_KINDS = ['NEW_GAME', 'REMIX_GAME', 'GAME_UPDATE', 'GAME_REVIEW', 'GAME_EVAL'];
+const PLAYDROP_PLUGIN_MANIFEST_REQUIRED_STARTUP_KINDS = WORKER_SUPPORTED_KINDS;
 const requireFromWorker = (0, node_module_1.createRequire)(__filename);
 const STAGED_PLAYDROP_PLUGIN_ROOT = '.playdrop/plugin';
 const PLAYDROP_PLUGIN_STAGING_MANIFEST_PATH = 'playdrop-worker-staging.json';
+const ASSIGNMENT_PLUGIN_BUNDLE_MAX_FILES = 1000;
+const ASSIGNMENT_PLUGIN_BUNDLE_MAX_UNCOMPRESSED_BYTES = 20 * 1024 * 1024;
 exports.WORKER_SESSION_EXPIRED_MESSAGE = 'worker session expired: run "playdrop auth login" and start the worker again';
 const WORKER_CONTEXT_ALLOWED_COMMANDS = [
     ['task', 'report'],
@@ -111,6 +117,7 @@ const WORKER_CONTEXT_ALLOWED_COMMANDS = [
     ['task', 'upload'],
     ['task', 'done'],
     ['task', 'submit-review'],
+    ['task', 'submit-eval'],
     ['task', 'fail'],
     ['review', 'validate-result'],
     ['review', 'compose-evidence'],
@@ -250,7 +257,7 @@ function normalizeAssignmentAgent(value) {
     return value;
 }
 function normalizeAssignmentKind(value, task) {
-    if (value !== 'NEW_GAME' && value !== 'REMIX_GAME' && value !== 'GAME_UPDATE' && value !== 'GAME_REVIEW') {
+    if (value !== 'NEW_GAME' && value !== 'REMIX_GAME' && value !== 'GAME_UPDATE' && value !== 'GAME_REVIEW' && value !== 'GAME_EVAL') {
         throw new Error('agent_task_assignment_kind_missing');
     }
     if (value !== task.kind) {
@@ -310,6 +317,9 @@ function resolveWorkerClaimTaskAssignment(claim) {
     if (!version) {
         throw new Error('agent_task_assignment_output_version_missing');
     }
+    const pluginBundle = assignment.pluginBundle && typeof assignment.pluginBundle === 'object' && !Array.isArray(assignment.pluginBundle)
+        ? assignment.pluginBundle
+        : null;
     return {
         taskId,
         taskToken,
@@ -329,10 +339,12 @@ function resolveWorkerClaimTaskAssignment(claim) {
             baseSource: inputs.baseSource ?? null,
             remixSource: inputs.remixSource ?? null,
             reviewTarget: inputs.reviewTarget ?? null,
+            evalTarget: inputs.evalTarget ?? null,
             allowedTemplateKeys: Array.isArray(inputs.allowedTemplateKeys)
                 ? Array.from(new Set(inputs.allowedTemplateKeys.map((value) => (typeof value === 'string' ? value.trim() : '')).filter(Boolean)))
                 : [],
         },
+        pluginBundle,
         output: {
             appId: typeof output.appId === 'number' && Number.isInteger(output.appId) && output.appId > 0 ? output.appId : null,
             appName: typeof output.appName === 'string' && output.appName.trim() ? output.appName.trim() : null,
@@ -503,7 +515,7 @@ function readTaskContextFile(startDir = node_process_1.default.cwd()) {
     if (!Number.isInteger(taskId) || taskId <= 0
         || !Number.isInteger(attempt) || attempt <= 0
         || !taskToken
-        || (parsed.kind !== 'NEW_GAME' && parsed.kind !== 'REMIX_GAME' && parsed.kind !== 'GAME_UPDATE' && parsed.kind !== 'GAME_REVIEW')
+        || (parsed.kind !== 'NEW_GAME' && parsed.kind !== 'REMIX_GAME' && parsed.kind !== 'GAME_UPDATE' && parsed.kind !== 'GAME_REVIEW' && parsed.kind !== 'GAME_EVAL')
         || (parsed.target !== 'FIRST_PARTY' && parsed.target !== 'PERSONAL')
         || !creatorUsername
         || !creatorRequest
@@ -883,8 +895,37 @@ function readPlaydropPluginStagingManifest(pluginRoot) {
     }
     const taskFiles = {};
     const rawTaskFiles = rawManifest.taskFiles;
+    for (const kind of PLAYDROP_PLUGIN_MANIFEST_REQUIRED_STARTUP_KINDS) {
+        const rawFiles = rawTaskFiles[kind];
+        if (!Array.isArray(rawFiles) || rawFiles.length <= 0) {
+            throw new Error(`playdrop_plugin_staging_manifest_incomplete:${kind}: non-empty taskFiles entry required in ${manifestPath}`);
+        }
+        const seen = new Set();
+        const files = rawFiles.map((value) => normalizePlaydropPluginManifestPath({ manifestPath, kind, value }));
+        for (const relativePath of files) {
+            if (seen.has(relativePath)) {
+                throw new Error(`playdrop_plugin_staging_manifest_duplicate_file:${kind}:${relativePath}`);
+            }
+            seen.add(relativePath);
+            const source = node_path_1.default.resolve(resolvedRoot, ...relativePath.split('/'));
+            const relativeToRoot = node_path_1.default.relative(resolvedRoot, source);
+            if (relativeToRoot.startsWith('..') || node_path_1.default.isAbsolute(relativeToRoot)) {
+                throw new Error(`playdrop_plugin_staging_manifest_invalid_path:${kind}:${relativePath}`);
+            }
+            if (!isFile(source)) {
+                throw new Error(`playdrop_plugin_staging_manifest_file_missing:${kind}:${relativePath}`);
+            }
+        }
+        taskFiles[kind] = files;
+    }
     for (const kind of WORKER_SUPPORTED_KINDS) {
+        if (taskFiles[kind]) {
+            continue;
+        }
         const rawFiles = rawTaskFiles[kind];
+        if (rawFiles === undefined) {
+            continue;
+        }
         if (!Array.isArray(rawFiles) || rawFiles.length <= 0) {
             throw new Error(`playdrop_plugin_staging_manifest_incomplete:${kind}: non-empty taskFiles entry required in ${manifestPath}`);
         }
@@ -908,7 +949,7 @@ function readPlaydropPluginStagingManifest(pluginRoot) {
     }
     return {
         version: 1,
-        taskFiles,
+        taskFiles: taskFiles,
     };
 }
 function resolvePlaydropPluginRoot(input = {}) {
@@ -945,6 +986,9 @@ async function stagePlaydropPluginReferences(input) {
     }
     const manifest = readPlaydropPluginStagingManifest(pluginRoot);
     const referenceFiles = manifest.taskFiles[input.kind];
+    if (!Array.isArray(referenceFiles) || referenceFiles.length <= 0) {
+        throw new Error(`playdrop_plugin_staging_manifest_incomplete:${input.kind}: non-empty taskFiles entry required`);
+    }
     const staged = [];
     for (const relativePath of referenceFiles) {
         const destinationRelativePath = node_path_1.default.join(STAGED_PLAYDROP_PLUGIN_ROOT, relativePath);
@@ -955,6 +999,54 @@ async function stagePlaydropPluginReferences(input) {
     }
     return staged;
 }
+async function stageAssignmentPluginBundle(input) {
+    const sha256 = typeof input.pluginBundle.sha256 === 'string' ? input.pluginBundle.sha256.trim() : '';
+    const contentBase64 = typeof input.pluginBundle.contentBase64 === 'string' ? input.pluginBundle.contentBase64.trim() : '';
+    if (!/^[a-f0-9]{64}$/.test(sha256) || !contentBase64) {
+        throw new Error('agent_task_plugin_bundle_invalid');
+    }
+    const buffer = Buffer.from(contentBase64, 'base64');
+    if (buffer.length <= 0) {
+        throw new Error('agent_task_plugin_bundle_empty');
+    }
+    const actualSha = (0, node_crypto_1.createHash)('sha256').update(buffer).digest('hex');
+    if (actualSha !== sha256) {
+        throw new Error('agent_task_plugin_bundle_sha_mismatch');
+    }
+    const pluginRoot = node_path_1.default.join(input.workspaceDir, '.playdrop', 'task-plugin');
+    await (0, promises_1.rm)(pluginRoot, { recursive: true, force: true });
+    await (0, promises_1.mkdir)(pluginRoot, { recursive: true });
+    const entries = (0, fflate_1.unzipSync)(new Uint8Array(buffer));
+    const entryPairs = Object.entries(entries);
+    const fileEntries = entryPairs.filter(([entryName]) => !entryName.replace(/\\/g, '/').endsWith('/'));
+    if (fileEntries.length > ASSIGNMENT_PLUGIN_BUNDLE_MAX_FILES) {
+        throw new Error(`agent_task_plugin_bundle_too_many_files:${fileEntries.length}>${ASSIGNMENT_PLUGIN_BUNDLE_MAX_FILES}`);
+    }
+    let uncompressedBytes = 0;
+    for (const [entryName, content] of Object.entries(entries)) {
+        const normalizedName = entryName.replace(/\\/g, '/');
+        if (!normalizedName || normalizedName.startsWith('/') || normalizedName.split('/').includes('..')) {
+            throw new Error(`agent_task_plugin_bundle_invalid_path:${entryName}`);
+        }
+        if (normalizedName.endsWith('/')) {
+            await (0, promises_1.mkdir)(node_path_1.default.join(pluginRoot, normalizedName), { recursive: true });
+            continue;
+        }
+        const destination = node_path_1.default.join(pluginRoot, normalizedName);
+        const relative = node_path_1.default.relative(pluginRoot, destination);
+        if (relative.startsWith('..') || node_path_1.default.isAbsolute(relative)) {
+            throw new Error(`agent_task_plugin_bundle_invalid_path:${entryName}`);
+        }
+        uncompressedBytes += content.byteLength;
+        if (uncompressedBytes > ASSIGNMENT_PLUGIN_BUNDLE_MAX_UNCOMPRESSED_BYTES) {
+            throw new Error(`agent_task_plugin_bundle_too_large:${uncompressedBytes}>${ASSIGNMENT_PLUGIN_BUNDLE_MAX_UNCOMPRESSED_BYTES}`);
+        }
+        await (0, promises_1.mkdir)(node_path_1.default.dirname(destination), { recursive: true });
+        await (0, promises_1.writeFile)(destination, Buffer.from(content));
+    }
+    readPlaydropPluginStagingManifest(pluginRoot);
+    return pluginRoot;
+}
 // A PlayDrop project root is the topmost directory holding a catalogue.json,
 // the same marker "playdrop project publish" resolves its target from. Nested
 // catalogue.json files belong to the root above them, so the walk stops at the
@@ -1316,6 +1408,7 @@ function buildWorkerCapabilities(input) {
         ? DEFAULT_WORKER_MAX_PARALLEL_TASKS
         : input.maxParallelTasks ?? DEFAULT_WORKER_MAX_PARALLEL_TASKS;
     const runningTaskCount = typeof input === 'string' ? 0 : input.runningTaskCount ?? 0;
+    const providerAccountLabel = typeof input === 'string' ? null : input.providerAccountLabel?.trim() || null;
     const degradedReasons = [];
     const agents = [];
     if (codexVersion) {
@@ -1368,6 +1461,7 @@ function buildWorkerCapabilities(input) {
         agents,
         ready: agents.some((agent) => agent.ready === true),
         degradedReasons,
+        ...(providerAccountLabel ? { providerAccountLabel } : {}),
         wrapperContractVersion: (0, runtime_1.readPositiveEnvInt)('PLAYDROP_WORKER_WRAPPER_CONTRACT_VERSION', WORKER_WRAPPER_CONTRACT_VERSION),
         ...(npmVersion ? { npmVersion } : {}),
         ...(playwright ? { playwright } : {}),
@@ -2170,6 +2264,7 @@ function collectWorkerLocalStatus() {
             playwright,
             runningTaskCount: 0,
             maxParallelTasks: (0, runtime_1.readPositiveEnvInt)('PLAYDROP_WORKER_MAX_PARALLEL_TASKS', DEFAULT_WORKER_MAX_PARALLEL_TASKS),
+            providerAccountLabel: node_process_1.default.env.PLAYDROP_WORKER_AGENT_ACCOUNT_LABEL?.trim() || null,
         });
         if (pluginVersion) {
             capabilities.pluginVersion = pluginVersion;
@@ -3031,6 +3126,7 @@ async function startWorker(options = {}) {
             throw new Error('worker_session_invalid: the stored session did not resolve to a user. Run "playdrop auth login" and retry.');
         }
         const target = resolveWorkerExecutionTargetFromRole(me.user.role);
+        const providerAccountLabel = readProviderAccountLabel(target);
         const codexVersion = probeCodexInstallation()?.codexVersion ?? null;
         const claudeVersion = probeClaudeInstallation()?.claudeVersion ?? null;
         const cursorVersion = probeCursorInstallation()?.cursorVersion ?? null;
@@ -3055,6 +3151,7 @@ async function startWorker(options = {}) {
             playwright,
             maxParallelTasks,
             runningTaskCount: 0,
+            providerAccountLabel,
         });
         if (playdropPluginVersion) {
             capabilities.pluginVersion = playdropPluginVersion;
@@ -3306,7 +3403,10 @@ async function startWorker(options = {}) {
                     message: 'Preparing worker workspace',
                     pct: 2,
                 });
-                availableSkillPaths = normalizeTelemetrySkillPaths(await stagePlaydropPluginReferences({ workspaceDir, pluginRoot: playdropPluginRoot, kind: task.kind }));
+                const assignmentPluginRoot = assignment.pluginBundle
+                    ? await stageAssignmentPluginBundle({ workspaceDir, pluginBundle: assignment.pluginBundle })
+                    : playdropPluginRoot;
+                availableSkillPaths = normalizeTelemetrySkillPaths(await stagePlaydropPluginReferences({ workspaceDir, pluginRoot: assignmentPluginRoot, kind: task.kind }));
                 await client.workerCreateAgentTaskEvent(task.id, {
                     workerKey,
                     leaseToken,
@@ -3417,6 +3517,14 @@ async function startWorker(options = {}) {
                         throw new Error('agent_task_assignment_review_target_missing');
                     }
                 }
+                else if (task.kind === 'GAME_EVAL') {
+                    if (assignment.inputs.baseSource !== null) {
+                        throw new Error('agent_task_assignment_unexpected_base_source');
+                    }
+                    if (!assignment.inputs.evalTarget) {
+                        throw new Error('agent_task_assignment_eval_target_missing');
+                    }
+                }
                 else {
                     throw new Error(`unsupported_agent_task_kind:${task.kind}`);
                 }
@@ -3454,7 +3562,7 @@ async function startWorker(options = {}) {
                         attempt: taskContext.attempt,
                         devPort,
                         claudeModel: claudeModel ?? resolveClaudeModel(assignment.model),
-                        playdropPluginRoot,
+                        playdropPluginRoot: assignmentPluginRoot,
                         onTranscriptChunks: async (chunks) => {
                             await appendTranscriptChunks(chunks);
                         },
@@ -3868,7 +3976,7 @@ async function resolveTaskCommandContext(command, optionsEnv, taskContext) {
 }
 async function uploadTask(options = {}) {
     const taskContext = readTaskContextFile();
-    if (taskContext.kind === 'GAME_REVIEW') {
+    if (taskContext.kind === 'GAME_REVIEW' || taskContext.kind === 'GAME_EVAL') {
         throw new Error('task_upload_not_allowed_for_game_review');
     }
     const workspaceDir = resolveTaskWorkspaceDir();
@@ -3965,9 +4073,13 @@ async function claimSlugTask(options) {
 }
 async function completeTask(options) {
     const taskContext = readTaskContextFile();
-    if (taskContext.kind === 'GAME_REVIEW') {
+    if (taskContext.kind === 'GAME_REVIEW' || taskContext.kind === 'GAME_EVAL') {
         throw new Error('task_done_not_allowed_for_game_review');
     }
+    const summary = typeof options.summary === 'string' ? options.summary.trim() : '';
+    if (!summary || summary.length > 200 || /[\r\n]/.test(summary)) {
+        throw new Error('task_done_summary_required');
+    }
     const ctx = await resolveTaskCommandContext('task done', options.env, taskContext);
     if (!ctx) {
         return;
@@ -4004,6 +4116,7 @@ async function completeTask(options) {
                 creatorUsername: uploadResult.creatorUsername,
                 completedBy: 'agent',
             },
+            summary,
             ...(nextSteps !== undefined ? { nextSteps } : {}),
         });
     }
@@ -4022,7 +4135,7 @@ async function completeTask(options) {
         }
         throw error;
     }
-    (0, output_1.printSuccess)('Task marked done. Send one short final status sentence now, then stop.');
+    (0, output_1.printSuccess)('Task marked done.');
 }
 function normalizeReviewSubmitState(value) {
     const normalized = typeof value === 'string' ? value.trim().toUpperCase() : '';
@@ -4171,6 +4284,29 @@ async function submitReviewTask(options) {
     });
     (0, output_1.printSuccess)('Review submitted. Send one short final status sentence now, then stop.');
 }
+async function submitEvalTask(options) {
+    const taskContext = readTaskContextFile();
+    if (taskContext.kind !== 'GAME_EVAL') {
+        throw new Error('task_submit_eval_requires_game_eval');
+    }
+    const resultFile = resolveReviewFilePath(options.resultFile, 'invalid_eval_result_file');
+    let result;
+    try {
+        result = JSON.parse((0, node_fs_1.readFileSync)(resultFile, 'utf8'));
+    }
+    catch (error) {
+        throw new Error(`invalid_eval_result_json:${error instanceof Error ? error.message : String(error)}`);
+    }
+    const ctx = await resolveTaskCommandContext('task submit-eval', options.env, taskContext);
+    if (!ctx) {
+        throw new Error('task_submit_eval_auth_required');
+    }
+    await ctx.client.workerSubmitAgentTaskEval(taskContext.taskId, {
+        taskToken: taskContext.taskToken,
+        result: result,
+    });
+    (0, output_1.printSuccess)('Eval submitted. Send one short final status sentence now, then stop.');
+}
 async function failTask(options) {
     const message = options.message?.trim();
     if (!message) {

package/dist/index.js

@@ -255,6 +255,7 @@ task
 task
     .command('done')
     .description('Mark the active agent task done after upload')
+    .requiredOption('--summary <summary>', 'One-sentence creator-facing recap of what was built')
     .option('--next-steps <path>', 'JSON file with up to three next-step suggestions')
     .option('--env <env>', 'Environment override')
     .action(async (opts) => {
@@ -272,6 +273,14 @@ task
     .action(async (opts) => {
     await (0, worker_1.submitReviewTask)(opts);
 });
+task
+    .command('submit-eval')
+    .description('Submit the active game eval task result')
+    .requiredOption('--result-file <path>', 'Strict GAME_EVAL JSON result file')
+    .option('--env <env>', 'Environment override')
+    .action(async (opts) => {
+    await (0, worker_1.submitEvalTask)(opts);
+});
 task
     .command('fail')
     .description('Fail the active agent task')
@@ -1040,6 +1049,7 @@ projectCapture
     .option('--expected-url <url>', 'Expected final URL after redirects')
     .option('--actions <path>', 'JSON action file with click, press, and wait steps before screenshot')
     .option('--viewport <size>', 'Viewport size as WIDTHxHEIGHT')
+    .option('--surface <surface>', 'Browser surface preset: desktop, mobile-landscape, or mobile-portrait')
     .option('--fps-sample-ms <milliseconds>', 'Sample frame rate for this many milliseconds after capture')
     .option('--username <username>', 'Login username')
     .option('--password <password>', 'Login password')

package/node_modules/@playdrop/api-client/dist/client.d.ts

@@ -1,4 +1,4 @@
-import type { AppleOAuthHandoffExchangeRequest, AppleOAuthHandoffExchangeResponse, AppleOAuthLinkStartRequest, AppleOAuthLinkStartResponse, ApplePushDeviceRequest, ApplePushDeviceStatusResponse, FirebasePushDeviceRequest, FirebasePushDeviceStatusResponse, ApiKeyLoginRequest, CliApiKeyStatusResponse, CliLoginApproveRequest, CliLoginApproveResponse, CliLoginPollRequest, CliLoginPollResponse, CliLoginStartResponse, CliLoginStatusResponse, CliWebLaunchStartRequest, CliWebLaunchStartResponse, CompleteXSignupRequest, CompleteXSignupResponse, DeleteMyAccountRequest, DeleteMyAccountResponse, LoginRequest, LoginResponse, LogoutResponse, PrivacyExportResponse, RegisterRequest, RegisterResponse, StarterProfileAssetsResponse, CreateAppRequest, CreateAppResponse, UpdateAppRequest, AdminUpdateAppRequest, UpdateAppResponse, AppsListResponse, AppBrowseRequest, AppDetailRequest, AppDetailResponse, AppMoreCollectionKey, AppMoreCollectionPage, AppMoreCollectionRequest, AppMoreCollectionsRequest, AppMoreCollectionsResponse, CreatorAppsBrowseRequest, AppRuntimeAssetsResponse, ContentLicense, InitializeAppUploadRequest, InitializeAppUploadResponse, UploadAppSessionFileResponse, UploadAppSessionOwnedAssetResponse, FinalizeAppUploadResponse, AppUploadLaunchCheckPreviewResponse, AbortAppUploadResponse, RecordAppUploadLaunchCheckRequest, RecordAppUploadLaunchCheckResponse, MyAppsResponse, BootstrapResponse, SetProfileAssetRequest, SetProfileAssetResponse, MeResponse, UserDetailResponse, UsersListResponse, CreateUserRequest, CreateUserResponse, UpdateAdminUserRequest, RemixScaffoldResponse, TemplateScaffoldResponse, HomeResponse, HomeCollectionPageResponse, FetchHomeOptions, DeleteAppResponse, AppType, AppAccessTokenResponse, DevAppAccessTokenResponse, DevPlayerResetResponse, AppLogEntryRequest, AppLogEntryResponse, ClaimFreeCreditRewardResponse, SubmitFeedbackRequest, SubmitFeedbackResponse, FeedbackListResponse, AdminFeedbackListResponse, AdminFeedbackResponse, UpdateAdminFeedbackRequest, UpdateAdminFeedbackResponse, DeleteFeedbackResponse, FeedbackCategory, FeedbackClient, FeedbackStatus, ShopResponse, AppleCreditPackFulfillmentRequest, AdInterstitialShowResponse, AdLoadResponse, AdShowRequest, AdRewardedShowResponse, AssetPurchaseRequest, BoostActivationRequest, BoostBalanceResponse, BoostHistoryResponse, BoostPurchaseRequest, BoostPurchaseResponse, BoostReportQuery, BoostReportResponse, BoostStatusResponse, AdLoadRequest, CreditPackCheckoutSessionRequest, PurchaseResponse, CreditPackCheckoutSessionResponse, CreditPackCheckoutStatusResponse, CreatorEarningsResponse, GoogleCreditPackFulfillmentRequest, MobileCreditPackFulfillmentResponse, TransactionsResponse, CreditPacksResponse, CreditPackResponse, PaymentProvider, PaymentStatus, AdminPaymentsListResponse, AdminPaymentDetailResponse, AdminPlatformAnalyticsRequest, AdminPlatformAnalyticsResponse, AdminRefundPaymentRequest, IapPurchaseRequest, IapPurchaseResponse, IapReceiptsResponse, IapReceiptListParams, IapReceiptResponse, AdminCreditTransactionsResponse, AiGenerationsResponse, AiGenerationDetailResponse, OwnedProfileAssetsResponse, SavedItemKind, CreateContentReportRequest, CreateCreatorReportRequest, CreateCommentReportRequest, CreatorBlockMutationResponse, LikedItemMutationResponse, DislikedItemMutationResponse, ModerationReportResponse, SavedItemMutationResponse, LibraryCategory, LibrarySort, LibraryResponse, TrackEngagementEventRequest, TrackEngagementEventResponse, CreatorFollowMutationResponse, ContentCommentsResponse, CreateContentCommentRequest, CreateContentCommentResponse, FollowingSort, FollowingCreatorsResponse, FollowingFeedSection, FollowingFeedResponse, DeleteContentCommentResponse, NotificationStatus, NotificationsResponse, NotificationReadResponse, EmailPreferencesResponse, UpdateUserCommunicationPreferencesRequest, PublicEmailPreferencesResponse, PublicEmailPreferencesUpdateRequest, EmailUnsubscribeResponse, AdminCreatorEngagementRulesResponse, AdminCreatorEngagementExecutionsResponse, AdminCreatorEngagementUserResponse, AdminCreatorEngagementRulePreviewRequest, AdminCreatorEngagementRulePreviewResponse, CreatorEngagementRuleExecutionStatus, PushSubscriptionMutationResponse, AdminUserResponse, DeleteAdminUserResponse, BatchCreateUsersResponse, UpsertWebPushSubscriptionRequest, UpdateProfileRequest, UpdateProfileResponse, WebPushConfigResponse, WebPushSubscriptionStatusResponse, ChangePasswordRequest, ChangePasswordResponse, UpgradeToCreatorResponse, AppleLinkRequest, AppleLinkResponse, AppleNativeChallengeResponse, ApplePendingSignupResponse, AppleSignInRequest, AppleSignInResponse, UnlinkXResponse, UnlinkGoogleResponse, GooglePendingSignupResponse, XPendingSignupResponse, GoogleOneTapExchangeRequest, GoogleOneTapExchangeResponse, UsernameAvailableResponse, CompleteAppleSignupRequest, CompleteAppleSignupResponse, CompleteGoogleSignupRequest, CompleteGoogleSignupResponse, GoogleAppleLinkRequest, GoogleAppleLinkResponse, GoogleAppleSignInRequest, GoogleAppleSignInResponse, AdminAppPlayerMetaResponse, AdminAppPlayerMetaUserResponse, AdminAppVersionReviewsListResponse, AdminEnqueueNextAppVersionReviewTaskResponse, AdminMutateAchievementStateRequest, AdminMutateAchievementStateResponse, AdminMutateLeaderboardScoreRequest, AdminMutateLeaderboardScoreResponse, AdminRetirePlayerMetaDefinitionResponse, AdminUpdateAppVersionReviewRequest, AdminUpdateAppVersionReviewResponse, CreateCliApiKeyResponse, GetAchievementResponse, GetLeaderboardResponse, ListAchievementsResponse, ListLeaderboardsResponse, MyAchievementsResponse, MyLeaderboardsResponse, SetAchievementProgressRequest, SetAchievementProgressResponse, SubmitLeaderboardScoreRequest, SubmitLeaderboardScoreResponse, UnlockAchievementResponse, ReviewState, AppVersionVisibility, AppVersionsListResponse, AppVersionDetailResponse, AppSourceBundleResponse, UpdateVersionResponse, DeleteVersionResponse, AdminAppsListResponse, AdminGameCollectionCandidatesResponse, AdminGameCollectionResponse, AdminGameCollectionsResponse, AdminUpdateAppResponse, CreateAdminGameCollectionRequest, DeleteAdminGameCollectionResponse, UpdateAdminGameCollectionGamesRequest, UpdateAdminGameCollectionRequest, AssetBrowseKind, AssetCategory, AssetFormat, AssetVisibility, AssetSourceKind, AssetListSort, AssetListResponse, AssetCategoriesResponse, AssetDetailResponse, AssetVersionsListResponse, CreateAssetVersionResponse, DeleteOwnedCustomAssetResponse, CreateAssetSpecVersionResponse, ListOwnedCustomAssetsResponse, LoadOwnedCustomAssetResponse, MutateOwnedCustomAssetRequest, MutateOwnedCustomAssetResponse, UpdateAssetRequest, UpdateAssetResponse, UpdateAssetVersionRequest, UpdateAssetVersionResponse, UpdateAssetSpecVersionResponse, DeleteAssetVersionResponse, DeleteAssetResponse, DeleteAssetSpecVersionResponse, DeleteAssetSpecResponse, AssetSourceBundleResponse, AssetPackListSort, AssetSpecAssetsListResponse, AssetSpecAppsListResponse, AssetSpecDetailResponse, AssetSpecListSort, AssetSpecListResponse, AssetSpecVersionResponseEnvelope, AssetSpecVersionsListResponse, AssetPackSourceBundleResponse, AssetPackListResponse, AssetPackDetailResponse, AssetPackVersionsListResponse, InitializeAssetPackUploadRequest, InitializeAssetPackUploadResponse, UploadAssetPackSessionAssetResponse, UploadAssetPackSessionMediaResponse, FinalizeAssetPackUploadResponse, AbortAssetPackUploadResponse, UpdateAssetPackRequest, UpdateAssetPackResponse, UpdateAssetPackVersionRequest, UpdateAssetPackVersionResponse, DeleteAssetPackVersionResponse, DeleteAssetPackResponse, BatchUpsertNodeRelationsRequest, BatchUpsertNodeRelationsResponse, ContentRemixRelationsResponse, SearchRequest, SearchResponse, SearchSuggestRequest, SearchSuggestResponse, AdminTagGroupResponse, AdminTagGroupsResponse, AdminTagResponse, CreateTagGroupRequest, CreateTagRequest, TagDirectoryResponse, TagDetailResponse, TagGroupDetailResponse, TagGroupSort, TagListResponse, UpdateTagGroupRequest, UpdateTagRequest, AdminAgentTaskDetailResponse, AdminAgentTaskTranscriptResponse, AdminAgentTasksResponse, AdminAgentWorkersStatusResponse, AgentTaskAvailabilityResponse, AgentTaskAttachmentUploadResponse, AgentTaskCreatorFeedbackStatus, AgentTaskDetailResponse, AgentTaskEventsResponse, AgentTaskKind, AgentTaskNewGameSuggestionsResponse, AgentTasksResponse, AppAgentTaskResponse, CancelAgentTaskResponse, ClearAgentTaskResponse, CreateAppAgentTaskRequest, CreateAppAgentTaskResponse, CreateNewGameAgentTaskRequest, CreateNewGameAgentTaskResponse, CreateRemixGameAgentTaskRequest, CreateRemixGameAgentTaskResponse, CreatorGameCreatorFeedbackStatusResponse, CreatorGamesListResponse, CreatorGameShareResponse, CreatorGameStatusResponse, CreatorGameVersionsResponse, MeAgentWorkersResponse, RetryAgentTaskResponse, WorkerClaimAgentTaskRequest, WorkerClaimAgentTaskResponse, WorkerClaimAgentTaskSlugRequest, WorkerClaimAgentTaskSlugResponse, WorkerCompleteAgentTaskRequest, WorkerCreateAgentTaskEventRequest, WorkerCreateAgentTaskEventResponse, WorkerCreateAgentTaskTranscriptChunkRequest, WorkerCreateAgentTaskTranscriptChunkResponse, WorkerAgentTaskAttemptUnavailableRequest, WorkerAgentTaskAttemptUnavailableResponse, WorkerFailAgentTaskRequest, WorkerHeartbeatAgentTaskRequest, WorkerHeartbeatAgentTaskResponse, WorkerRecordAgentTaskRunTelemetryRequest, WorkerRecordAgentTaskRunTelemetryResponse, WorkerPresenceRequest, WorkerPresenceResponse, WorkerSubmitAgentTaskReviewRequest, WorkerSubmitAgentTaskReviewResponse, WorkerUpdateIntentRequest, WorkerUpdateIntentResponse, FreeCreditsResponse, InvitePreviewResponse, MyInviteCodeResponse } from '@playdrop/types';
+import type { AppleOAuthHandoffExchangeRequest, AppleOAuthHandoffExchangeResponse, AppleOAuthLinkStartRequest, AppleOAuthLinkStartResponse, ApplePushDeviceRequest, ApplePushDeviceStatusResponse, FirebasePushDeviceRequest, FirebasePushDeviceStatusResponse, ApiKeyLoginRequest, CliApiKeyStatusResponse, CliLoginApproveRequest, CliLoginApproveResponse, CliLoginPollRequest, CliLoginPollResponse, CliLoginStartResponse, CliLoginStatusResponse, CliWebLaunchStartRequest, CliWebLaunchStartResponse, CompleteXSignupRequest, CompleteXSignupResponse, DeleteMyAccountRequest, DeleteMyAccountResponse, LoginRequest, LoginResponse, LogoutResponse, PrivacyExportResponse, RegisterRequest, RegisterResponse, StarterProfileAssetsResponse, CreateAppRequest, CreateAppResponse, UpdateAppRequest, AdminUpdateAppRequest, UpdateAppResponse, AppsListResponse, AppBrowseRequest, AppDetailRequest, AppDetailResponse, AppMoreCollectionKey, AppMoreCollectionPage, AppMoreCollectionRequest, AppMoreCollectionsRequest, AppMoreCollectionsResponse, CreatorAppsBrowseRequest, AppRuntimeAssetsResponse, ContentLicense, InitializeAppUploadRequest, InitializeAppUploadResponse, UploadAppSessionFileResponse, UploadAppSessionOwnedAssetResponse, FinalizeAppUploadResponse, AppUploadLaunchCheckPreviewResponse, AbortAppUploadResponse, RecordAppUploadLaunchCheckRequest, RecordAppUploadLaunchCheckResponse, MyAppsResponse, BootstrapResponse, SetProfileAssetRequest, SetProfileAssetResponse, MeResponse, UserDetailResponse, UsersListResponse, CreateUserRequest, CreateUserResponse, UpdateAdminUserRequest, RemixScaffoldResponse, TemplateScaffoldResponse, HomeResponse, HomeCollectionPageResponse, FetchHomeOptions, DeleteAppResponse, AppType, AppAccessTokenResponse, DevAppAccessTokenResponse, DevPlayerResetResponse, AppLogEntryRequest, AppLogEntryResponse, ClaimFreeCreditRewardResponse, SubmitFeedbackRequest, SubmitFeedbackResponse, FeedbackListResponse, AdminFeedbackListResponse, AdminFeedbackResponse, UpdateAdminFeedbackRequest, UpdateAdminFeedbackResponse, DeleteFeedbackResponse, FeedbackCategory, FeedbackClient, FeedbackStatus, ShopResponse, AppleCreditPackFulfillmentRequest, AdInterstitialShowResponse, AdLoadResponse, AdShowRequest, AdRewardedShowResponse, AssetPurchaseRequest, BoostActivationRequest, BoostBalanceResponse, BoostHistoryResponse, BoostPurchaseRequest, BoostPurchaseResponse, BoostReportQuery, BoostReportResponse, BoostStatusResponse, AdLoadRequest, CreditPackCheckoutSessionRequest, PurchaseResponse, CreditPackCheckoutSessionResponse, CreditPackCheckoutStatusResponse, CreatorEarningsResponse, GoogleCreditPackFulfillmentRequest, MobileCreditPackFulfillmentResponse, TransactionsResponse, CreditPacksResponse, CreditPackResponse, PaymentProvider, PaymentStatus, AdminPaymentsListResponse, AdminPaymentDetailResponse, AdminPlatformAnalyticsRequest, AdminPlatformAnalyticsResponse, AdminRefundPaymentRequest, IapPurchaseRequest, IapPurchaseResponse, IapReceiptsResponse, IapReceiptListParams, IapReceiptResponse, AdminCreditTransactionsResponse, AiGenerationsResponse, AiGenerationDetailResponse, OwnedProfileAssetsResponse, SavedItemKind, CreateContentReportRequest, CreateCreatorReportRequest, CreateCommentReportRequest, CreatorBlockMutationResponse, LikedItemMutationResponse, DislikedItemMutationResponse, ModerationReportResponse, SavedItemMutationResponse, LibraryCategory, LibrarySort, LibraryResponse, TrackEngagementEventRequest, TrackEngagementEventResponse, CreatorFollowMutationResponse, ContentCommentsResponse, CreateContentCommentRequest, CreateContentCommentResponse, FollowingSort, FollowingCreatorsResponse, FollowingFeedSection, FollowingFeedResponse, DeleteContentCommentResponse, NotificationStatus, NotificationsResponse, NotificationReadResponse, EmailPreferencesResponse, UpdateUserCommunicationPreferencesRequest, PublicEmailPreferencesResponse, PublicEmailPreferencesUpdateRequest, EmailUnsubscribeResponse, AdminCreatorEngagementRulesResponse, AdminCreatorEngagementExecutionsResponse, AdminCreatorEngagementUserResponse, AdminCreatorEngagementRulePreviewRequest, AdminCreatorEngagementRulePreviewResponse, CreatorEngagementRuleExecutionStatus, PushSubscriptionMutationResponse, AdminUserResponse, DeleteAdminUserResponse, BatchCreateUsersResponse, UpsertWebPushSubscriptionRequest, UpdateProfileRequest, UpdateProfileResponse, WebPushConfigResponse, WebPushSubscriptionStatusResponse, ChangePasswordRequest, ChangePasswordResponse, UpgradeToCreatorResponse, AppleLinkRequest, AppleLinkResponse, AppleNativeChallengeResponse, ApplePendingSignupResponse, AppleSignInRequest, AppleSignInResponse, UnlinkXResponse, UnlinkGoogleResponse, GooglePendingSignupResponse, XPendingSignupResponse, GoogleOneTapExchangeRequest, GoogleOneTapExchangeResponse, UsernameAvailableResponse, CompleteAppleSignupRequest, CompleteAppleSignupResponse, CompleteGoogleSignupRequest, CompleteGoogleSignupResponse, GoogleAppleLinkRequest, GoogleAppleLinkResponse, GoogleAppleSignInRequest, GoogleAppleSignInResponse, AdminAppPlayerMetaResponse, AdminAppPlayerMetaUserResponse, AdminAppVersionReviewsListResponse, AdminEnqueueNextAppVersionReviewTaskResponse, AdminMutateAchievementStateRequest, AdminMutateAchievementStateResponse, AdminMutateLeaderboardScoreRequest, AdminMutateLeaderboardScoreResponse, AdminRetirePlayerMetaDefinitionResponse, AdminUpdateAppVersionReviewRequest, AdminUpdateAppVersionReviewResponse, CreateCliApiKeyResponse, GetAchievementResponse, GetLeaderboardResponse, ListAchievementsResponse, ListLeaderboardsResponse, MyAchievementsResponse, MyLeaderboardsResponse, SetAchievementProgressRequest, SetAchievementProgressResponse, SubmitLeaderboardScoreRequest, SubmitLeaderboardScoreResponse, UnlockAchievementResponse, ReviewState, AppVersionVisibility, AppVersionsListResponse, AppVersionDetailResponse, AppSourceBundleResponse, UpdateVersionResponse, DeleteVersionResponse, AdminAppsListResponse, AdminGameCollectionCandidatesResponse, AdminGameCollectionResponse, AdminGameCollectionsResponse, AdminUpdateAppResponse, CreateAdminGameCollectionRequest, DeleteAdminGameCollectionResponse, UpdateAdminGameCollectionGamesRequest, UpdateAdminGameCollectionRequest, AssetBrowseKind, AssetCategory, AssetFormat, AssetVisibility, AssetSourceKind, AssetListSort, AssetListResponse, AssetCategoriesResponse, AssetDetailResponse, AssetVersionsListResponse, CreateAssetVersionResponse, DeleteOwnedCustomAssetResponse, CreateAssetSpecVersionResponse, ListOwnedCustomAssetsResponse, LoadOwnedCustomAssetResponse, MutateOwnedCustomAssetRequest, MutateOwnedCustomAssetResponse, UpdateAssetRequest, UpdateAssetResponse, UpdateAssetVersionRequest, UpdateAssetVersionResponse, UpdateAssetSpecVersionResponse, DeleteAssetVersionResponse, DeleteAssetResponse, DeleteAssetSpecVersionResponse, DeleteAssetSpecResponse, AssetSourceBundleResponse, AssetPackListSort, AssetSpecAssetsListResponse, AssetSpecAppsListResponse, AssetSpecDetailResponse, AssetSpecListSort, AssetSpecListResponse, AssetSpecVersionResponseEnvelope, AssetSpecVersionsListResponse, AssetPackSourceBundleResponse, AssetPackListResponse, AssetPackDetailResponse, AssetPackVersionsListResponse, InitializeAssetPackUploadRequest, InitializeAssetPackUploadResponse, UploadAssetPackSessionAssetResponse, UploadAssetPackSessionMediaResponse, FinalizeAssetPackUploadResponse, AbortAssetPackUploadResponse, UpdateAssetPackRequest, UpdateAssetPackResponse, UpdateAssetPackVersionRequest, UpdateAssetPackVersionResponse, DeleteAssetPackVersionResponse, DeleteAssetPackResponse, BatchUpsertNodeRelationsRequest, BatchUpsertNodeRelationsResponse, ContentRemixRelationsResponse, SearchRequest, SearchResponse, SearchSuggestRequest, SearchSuggestResponse, AdminTagGroupResponse, AdminTagGroupsResponse, AdminTagResponse, CreateTagGroupRequest, CreateTagRequest, TagDirectoryResponse, TagDetailResponse, TagGroupDetailResponse, TagGroupSort, TagListResponse, UpdateTagGroupRequest, UpdateTagRequest, AdminAgentTaskDetailResponse, AdminAgentTaskTranscriptResponse, AdminAgentTasksResponse, AdminAgentWorkersStatusResponse, AgentTaskAvailabilityResponse, AgentTaskAttachmentUploadResponse, AgentTaskCreatorFeedbackStatus, AgentTaskDetailResponse, AgentTaskEventsResponse, AgentTaskKind, AgentTaskNewGameSuggestionsResponse, AgentTasksResponse, AppAgentTaskResponse, CancelAgentTaskResponse, ClearAgentTaskResponse, CreateAppAgentTaskRequest, CreateAppAgentTaskResponse, CreateGameEvalAgentTaskRequest, CreateGameEvalAgentTaskResponse, CreateNewGameAgentTaskRequest, CreateNewGameAgentTaskResponse, CreateRemixGameAgentTaskRequest, CreateRemixGameAgentTaskResponse, CreatorGameCreatorFeedbackStatusResponse, CreatorGamesListResponse, CreatorGameShareResponse, CreatorGameStatusResponse, CreatorGameVersionsResponse, MeAgentWorkersResponse, RetryAgentTaskResponse, WorkerClaimAgentTaskRequest, WorkerClaimAgentTaskResponse, WorkerClaimAgentTaskSlugRequest, WorkerClaimAgentTaskSlugResponse, WorkerCompleteAgentTaskRequest, WorkerCreateAgentTaskEventRequest, WorkerCreateAgentTaskEventResponse, WorkerCreateAgentTaskTranscriptChunkRequest, WorkerCreateAgentTaskTranscriptChunkResponse, WorkerAgentTaskAttemptUnavailableRequest, WorkerAgentTaskAttemptUnavailableResponse, WorkerFailAgentTaskRequest, WorkerHeartbeatAgentTaskRequest, WorkerHeartbeatAgentTaskResponse, WorkerRecordAgentTaskRunTelemetryRequest, WorkerRecordAgentTaskRunTelemetryResponse, WorkerPresenceRequest, WorkerPresenceResponse, WorkerSubmitAgentTaskEvalRequest, WorkerSubmitAgentTaskEvalResponse, WorkerSubmitAgentTaskReviewRequest, WorkerSubmitAgentTaskReviewResponse, WorkerUpdateIntentRequest, WorkerUpdateIntentResponse, FreeCreditsResponse, InvitePreviewResponse, MyInviteCodeResponse } from '@playdrop/types';
 import { type AdminAgentTaskTranscriptOptions, type ListCreatorGamesOptions, type ListAgentTasksOptions } from './domains/agent-tasks.js';
 export interface HttpMethod {
     GET: 'GET';
@@ -135,6 +135,7 @@ export interface ApiClient {
     uploadAgentTaskAttachment(file: Blob, description?: string, filename?: string): Promise<AgentTaskAttachmentUploadResponse>;
     createNewGameAgentTask(body: CreateNewGameAgentTaskRequest): Promise<CreateNewGameAgentTaskResponse>;
     createRemixGameAgentTask(body: CreateRemixGameAgentTaskRequest): Promise<CreateRemixGameAgentTaskResponse>;
+    createGameEvalAgentTask(body: CreateGameEvalAgentTaskRequest): Promise<CreateGameEvalAgentTaskResponse>;
     createAppAgentTask(appName: string, body: CreateAppAgentTaskRequest): Promise<CreateAppAgentTaskResponse>;
     getAppAgentTask(appName: string): Promise<AppAgentTaskResponse>;
     getAppAgentTaskAvailability(appName: string): Promise<AgentTaskAvailabilityResponse>;
@@ -163,6 +164,7 @@ export interface ApiClient {
     workerRecordAgentTaskRunTelemetry(taskId: number, body: WorkerRecordAgentTaskRunTelemetryRequest): Promise<WorkerRecordAgentTaskRunTelemetryResponse>;
     workerCompleteAgentTask(taskId: number, body: WorkerCompleteAgentTaskRequest): Promise<AgentTaskDetailResponse>;
     workerSubmitAgentTaskReview(taskId: number, body: WorkerSubmitAgentTaskReviewRequest): Promise<WorkerSubmitAgentTaskReviewResponse>;
+    workerSubmitAgentTaskEval(taskId: number, body: WorkerSubmitAgentTaskEvalRequest): Promise<WorkerSubmitAgentTaskEvalResponse>;
     workerFailAgentTask(taskId: number, body: WorkerFailAgentTaskRequest): Promise<AgentTaskDetailResponse>;
     adminListAgentTasks(options?: ListAgentTasksOptions): Promise<AdminAgentTasksResponse>;
     adminGetAgentTask(taskId: number): Promise<AdminAgentTaskDetailResponse>;