dominds 1.17.7 → 1.18.2

package/dist/runtime/driver-messages.d.ts

@@ -9,6 +9,7 @@ export declare function formatNewCourseStartPrompt(language: LanguageCode, args:
     source: 'clear_mind' | 'critical_auto_clear';
 }): string;
 export declare function formatDiligenceAutoContinuePrompt(language: LanguageCode, diligenceText: string): string;
+export declare function formatReminderContextGuide(language: LanguageCode): string;
 export declare function formatReminderItemGuide(language: LanguageCode, reminderId: string, content: string, options?: {
     meta?: unknown;
     scope?: 'dialog' | 'personal' | 'agent_shared';

package/dist/runtime/driver-messages.js

@@ -7,6 +7,7 @@ exports.formatRegisteredTellaskCallerUpdateNotice = formatRegisteredTellaskCalle
 exports.formatRegisteredTellaskCalleeUpdateNotice = formatRegisteredTellaskCalleeUpdateNotice;
 exports.formatNewCourseStartPrompt = formatNewCourseStartPrompt;
 exports.formatDiligenceAutoContinuePrompt = formatDiligenceAutoContinuePrompt;
+exports.formatReminderContextGuide = formatReminderContextGuide;
 exports.formatReminderItemGuide = formatReminderItemGuide;
 exports.formatQ4HDiligencePushBudgetExhausted = formatQ4HDiligencePushBudgetExhausted;
 exports.formatDomindsNoteTellaskForTeammatesOnly = formatDomindsNoteTellaskForTeammatesOnly;
@@ -149,6 +150,25 @@ function formatDiligenceAutoContinuePrompt(language, diligenceText) {
         trimmed,
     ].join('\n');
 }
+function formatReminderContextGuide(language) {
+    if (language === 'zh') {
+        return [
+            formatSystemNoticePrefix(language),
+            '以下是当前可见提醒项的运行时上下文投影。它们会按各自语义进入模型上下文（例如 self-reminder 保持 assistant-side 工作集语义），但不是我刚刚生成给用户的聊天正文。',
+            '在 WebUI 中，用户通过独立的 Reminder 小组件/面板项看到这些提醒，并能把它们和聊天正文区分开。',
+            '提醒项只作为工作集/状态参考；只有实际改变当前判断、计划或风险的信息，才需要提炼进后续有实质内容的对外回复。',
+        ].join('\n');
+    }
+    return [
+        formatSystemNoticePrefix(language),
+        'The following visible reminders are runtime-added context projections. They enter model context according to their own semantics (for example, self-reminders keep assistant-side workset semantics), but they are not chat text I just generated for the user.',
+        'In the WebUI, the user sees these reminders through a separate Reminder widget/panel item and can distinguish them from the chat transcript.',
+        'Use reminders as workset/state references; only carry information into a later substantive outward reply when it materially changes current judgment, plan, or risk.',
+    ].join('\n');
+}
+function formatReminderItemProjectionNote(language) {
+    return language === 'zh' ? 'Reminder 上下文投影条目：' : 'Reminder context projection item:';
+}
 function formatReminderItemGuide(language, reminderId, content, options) {
     function isRecord(value) {
         return typeof value === 'object' && value !== null && !Array.isArray(value);
@@ -190,13 +210,15 @@ function formatReminderItemGuide(language, reminderId, content, options) {
             : isPendingTellaskReminder && pendingTellaskCount === 0
                 ? `If I have confirmed this is only noise cleanup and not an action step, I may run: delete_reminder({ "reminder_id": "${reminderId}" })`
                 : `If I need to delete this reminder, run: delete_reminder({ "reminder_id": "${reminderId}" })`;
+    const projectionNote = formatReminderItemProjectionNote(language);
+    const enProjectionPrefix = `${projectionNote} `;
     if (language === 'zh') {
         if (managementTool) {
             const updateInstructionSafe = updateInstruction ?? `${managementTool}({ ... })`;
             return [
                 `提醒项 [${reminderId}]（工具状态）`,
                 '',
-                '我把这条当作工具维护的状态参考。默认不在对外回复里专门确认、复述或总结它；只有它实际改变当前判断、计划或风险时，我才提炼真正相关的部分。',
+                `${projectionNote}我把这条当作工具维护的状态参考。默认不在对外回复里专门确认、复述或总结它；只有它实际改变当前判断、计划或风险时，我才提炼真正相关的部分。`,
                 '',
                 `这条提醒项由工具 ${managementTool} 管理；如果我要调整它，就用 ${managementTool}（不要用 update_reminder）。`,
                 '',
@@ -211,7 +233,7 @@ function formatReminderItemGuide(language, reminderId, content, options) {
             return [
                 `提醒项 [${reminderId}]`,
                 '',
-                '这是带有 meta 控制更新规则的提醒项。我仍把它当作状态参考，但不要用 update_reminder 直接改写内容。',
+                `${projectionNote}这是带有 meta 控制更新规则的提醒项。我仍把它当作状态参考，但不要用 update_reminder 直接改写内容。`,
                 '',
                 `如果我要更新这条提醒项，不能用 update_reminder；请按此处理：${updateInstruction}`,
                 deleteInstruction,
@@ -224,7 +246,7 @@ function formatReminderItemGuide(language, reminderId, content, options) {
             return [
                 `提醒项 [${reminderId}]（换程接续信息）`,
                 '',
-                '我把这条当作换程后快速恢复工作的接续包，不把它自动当成当前必须立刻执行的指令。',
+                `${projectionNote}我把这条当作换程后快速恢复工作的接续包，不把它自动当成当前必须立刻执行的指令。`,
                 '',
                 '我应优先保留下一步行动、关键定位、运行/验证信息、容易丢的临时细节；不要重复差遣牒已覆盖的内容。进入新一程后，我的第一步就是以清醒头脑重新审视并整理更新：删除冗余、纠正偏激/失真思路、压缩成高质量提醒项。若目前只是粗略过桥笔记，进入新一程后我必须尽快收敛。',
                 '',
@@ -239,8 +261,8 @@ function formatReminderItemGuide(language, reminderId, content, options) {
             `提醒项 [${reminderId}]${scope === 'personal' ? '（个人范围）' : ''}`,
             '',
             scope === 'personal'
-                ? '这是我给自己的个人范围显眼提示；在所有由我主理的后续对话里都会看到它。我不把它自动当成系统下发的下一步动作。'
-                : '这是我给自己的显眼提示，用于保留当前对话里容易丢的工作信息；我不把它自动当成系统下发的下一步动作。',
+                ? `${projectionNote}这是我给自己的个人范围显眼提示；在所有由我主理的后续对话里都会看到它。我不把它自动当成系统下发的下一步动作。`
+                : `${projectionNote}这是我给自己的显眼提示，用于保留当前对话里容易丢的工作信息；我不把它自动当成系统下发的下一步动作。`,
             '',
             scope === 'personal'
                 ? '我应保持简洁、及时更新；不再需要时就删除。若它只对当前对话有效，应改写成 dialog 范围提醒而不是长期堆在个人范围里。'
@@ -257,7 +279,7 @@ function formatReminderItemGuide(language, reminderId, content, options) {
         const updateInstructionSafe = updateInstruction ?? `${managementTool}({ ... })`;
         return `REMINDER [${reminderId}] (TOOL STATE)
 
-I treat this as a tool-maintained state reference. By default I should not explicitly acknowledge, restate, or summarize it in my outward reply; I should only extract the parts that materially change my current judgment, plan, or risk.
+${enProjectionPrefix}I treat this as a tool-maintained state reference. By default I should not explicitly acknowledge, restate, or summarize it in my outward reply; I should only extract the parts that materially change my current judgment, plan, or risk.
 
 This reminder is managed by tool ${managementTool}; if I need to change it, I should use ${managementTool} instead of update_reminder.
 
@@ -269,7 +291,7 @@ ${content}`;
     if (updateInstruction) {
         return `REMINDER [${reminderId}]
 
-This reminder has a meta-controlled update path. I should still treat it as state/reference, and I must not rewrite it directly with update_reminder.
+${enProjectionPrefix}This reminder has a meta-controlled update path. I should still treat it as state/reference, and I must not rewrite it directly with update_reminder.
 
 If I need to update this reminder, I must not use update_reminder; follow instead: ${updateInstruction}
 ${deleteInstruction}
@@ -279,7 +301,7 @@ ${content}`;
     if (isContinuationPackageReminder) {
         return `REMINDER [${reminderId}] (CONTINUATION PACKAGE)
 
-I treat this as resume information for the next course, not as an automatic must-do command.
+${enProjectionPrefix}I treat this as resume information for the next course, not as an automatic must-do command.
 
 I should keep the next step, key pointers, run/verify info, and easy-to-lose volatile details here. I should not duplicate Taskdoc content. In the new course, my first step is to review and rewrite this with a clear head: remove redundancy, correct biased or distorted bridge notes, and compress it into a high-quality reminder. If this is only a rough bridge note, I should reconcile it early in the new course.
 
@@ -291,8 +313,8 @@ ${content}`;
     return `REMINDER [${reminderId}]${scope === 'personal' ? ' (PERSONAL SCOPE)' : ''}
 
 ${scope === 'personal'
-        ? 'This is my conspicuous personal-scope reminder. I will keep seeing it in all later dialogs I lead, and I do not treat it as an automatically assigned next action.'
-        : 'This is my conspicuous self-reminder for easy-to-lose work details in the current dialog. I do not treat it as an automatically assigned next action.'}
+        ? `${enProjectionPrefix}This is my conspicuous personal-scope reminder. I will keep seeing it in all later dialogs I lead, and I do not treat it as an automatically assigned next action.`
+        : `${enProjectionPrefix}This is my conspicuous self-reminder for easy-to-lose work details in the current dialog. I do not treat it as an automatically assigned next action.`}
 
 ${scope === 'personal'
         ? 'I should keep it concise, refresh it when needed, and delete it when obsolete. If it is only useful for the current dialog, I should rewrite it into dialog scope instead of letting personal scope accumulate noise.'

package/dist/runtime/reply-prompt-copy.js

@@ -68,8 +68,8 @@ function buildSubdialogRoleHeaderCopy(args) {
     const requester = `@${args.requesterId}`;
     if (args.expectedReplyTool === undefined) {
         return args.language === 'zh'
-            ? `${requester} 已通过诉请安排你处理下述诉请内容。只有需要回问时，才调用 \`tellaskBack\`。`
-            : `${requester} has assigned you to handle the request content below. Call \`tellaskBack\` only when you truly need to ask back.`;
+            ? `${requester} 已通过诉请安排你处理下述诉请内容。只有确实需要向上游回问、且现有规程无法直接判责时，才调用 \`tellaskBack\`。`
+            : `${requester} has assigned you to handle the request content below. Call \`tellaskBack\` only when you truly need to ask upstream back and existing SOP cannot directly identify another owner.`;
     }
     const kindLabel = (0, tellask_labels_1.getTellaskKindLabel)({
         language: args.language,
@@ -77,8 +77,8 @@ function buildSubdialogRoleHeaderCopy(args) {
         bracketed: true,
     });
     return args.language === 'zh'
-        ? `${requester} 已通过${kindLabel}安排你处理下述诉请内容。等你准备好回复内容后，调用 \`${args.expectedReplyTool}\` 完成回复。只有需要回问时，才调用 \`tellaskBack\`。`
-        : `${requester} has assigned you, via this ${kindLabel}, to handle the request content below. Once your reply content is ready, call \`${args.expectedReplyTool}\` to deliver it. Call \`tellaskBack\` only when you truly need to ask back.`;
+        ? `${requester} 已通过${kindLabel}安排你处理下述诉请内容。等你准备好回复内容后，调用 \`${args.expectedReplyTool}\` 完成回复。只有确实需要向上游回问、且现有规程无法直接判责时，才调用 \`tellaskBack\`。`
+        : `${requester} has assigned you, via this ${kindLabel}, to handle the request content below. Once your reply content is ready, call \`${args.expectedReplyTool}\` to deliver it. Call \`tellaskBack\` only when you truly need to ask upstream back and existing SOP cannot directly identify another owner.`;
 }
 function buildReplyObligationSuppressionGuideText(language) {
     if (language === 'zh') {

package/dist/server/api-routes.js

@@ -2109,6 +2109,13 @@ async function handleGetTeamConfig(res) {
 async function handleGetDialogs(res, status) {
     try {
         const rootDialogs = [];
+        const allDialogIds = await persistence_1.DialogPersistence.listAllDialogIds(status);
+        const subdialogCountByRootId = new Map();
+        for (const dialogId of allDialogIds) {
+            if (dialogId.selfId === dialogId.rootId)
+                continue;
+            subdialogCountByRootId.set(dialogId.rootId, (subdialogCountByRootId.get(dialogId.rootId) ?? 0) + 1);
+        }
         const ids = await persistence_1.DialogPersistence.listDialogs(status);
         for (const id of ids) {
             const dialogId = new dialog_1.DialogID(id);
@@ -2124,9 +2131,7 @@ async function handleGetDialogs(res, status) {
                 continue;
             }
             const waitingForFreshBootsReasoning = await detectWaitingForFreshBootsReasoning(dialogId, status);
-            // Count subdialogs for this root dialog
-            const subPath = path.join(rootPath, 'subdialogs');
-            const subdialogCount = await countSubdialogs(subPath);
+            const subdialogCount = subdialogCountByRootId.get(meta.id) ?? 0;
             if (!(await pathStillExistsForLookup(rootPath))) {
                 continue;
             }
@@ -2170,45 +2175,6 @@ async function handleGetRunControlCounts(res) {
         return true;
     }
 }
-/**
- * Count subdialog directories recursively
- */
-async function countSubdialogs(dirPath) {
-    try {
-        const entries = await promises_1.default.readdir(dirPath, { withFileTypes: true });
-        let count = 0;
-        for (const entry of entries) {
-            if (entry.isDirectory()) {
-                const fullPath = path.join(dirPath, entry.name);
-                const dialogYamlPath = path.join(fullPath, 'dialog.yaml');
-                try {
-                    await promises_1.default.access(dialogYamlPath);
-                    // This directory contains dialog.yaml - it's a subdialog.
-                    count++;
-                }
-                catch (error) {
-                    if (getErrorCode(error) === 'ENOENT') {
-                        // Only a missing dialog.yaml means "keep recursing". Permission/I/O failures must stay
-                        // loud so operators can distinguish inaccessible paths from ordinary non-dialog dirs.
-                        count += await countSubdialogs(fullPath);
-                        continue;
-                    }
-                    throw error;
-                }
-            }
-        }
-        return count;
-    }
-    catch (error) {
-        if (typeof error === 'object' &&
-            error !== null &&
-            'code' in error &&
-            error.code === 'ENOENT') {
-            return 0;
-        }
-        throw error;
-    }
-}
 /**
  * Get full hierarchy (root + subdialogs) for a single root dialog
  */
@@ -2292,7 +2258,7 @@ async function handleGetDialogHierarchy(res, rootId, status) {
         respondJson(res, 200, {
             success: true,
             hierarchy: {
-                root: rootInfo,
+                root: { ...rootInfo, subdialogCount: subdialogs.length },
                 subdialogs,
             },
         });
@@ -2331,6 +2297,7 @@ async function handleGetDialogListSubdialogNode(res, dialog) {
         }
         const latest = await loadDialogLatestForLookup(dialogId, status, 'handleGetDialogListSubdialogNode');
         const waitingForFreshBootsReasoning = await detectWaitingForFreshBootsReasoning(dialogId, status);
+        const rootSubdialogCount = await persistence_1.DialogPersistence.countAllSubdialogsUnderRoot(new dialog_1.DialogID(dialog.rootId), status);
         const subdialogPath = persistence_1.DialogPersistence.getSubdialogPath(dialogId, status);
         if (!(await pathStillExistsForLookup(subdialogPath))) {
             respondJson(res, 404, { success: false, error: `Dialog not found in ${status}` });
@@ -2347,6 +2314,7 @@ async function handleGetDialogListSubdialogNode(res, dialog) {
             subdialogNode: {
                 selfId: metadata.id,
                 rootId: dialog.rootId,
+                rootSubdialogCount,
                 supdialogId: derivedSupdialogId,
                 agentId: metadata.agentId,
                 taskDocPath: metadata.taskDocPath,

package/dist/server/websocket-handler.js

@@ -1,4 +1,7 @@
 "use strict";
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.shouldQueueUserSupplementAtGenerationBoundary = shouldQueueUserSupplementAtGenerationBoundary;
 exports.handleWebSocketMessage = handleWebSocketMessage;
@@ -14,6 +17,9 @@ const evt_1 = require("@longrun-ai/kernel/evt");
 const language_1 = require("@longrun-ai/kernel/types/language");
 const storage_1 = require("@longrun-ai/kernel/types/storage");
 const time_1 = require("@longrun-ai/kernel/utils/time");
+const crypto_1 = require("crypto");
+const promises_1 = __importDefault(require("fs/promises"));
+const path_1 = __importDefault(require("path"));
 const ws_1 = require("ws");
 const runtime_1 = require("../apps/runtime");
 const global_dialog_event_broadcaster_1 = require("../bootstrap/global-dialog-event-broadcaster");
@@ -54,6 +60,113 @@ function normalizeDiligencePushMax(value) {
         return 0;
     return Math.floor(value);
 }
+const USER_IMAGE_ATTACHMENT_MAX_COUNT = 10;
+const USER_IMAGE_ATTACHMENT_MAX_BYTES = 10 * 1024 * 1024;
+function userImageMimeTypeToExt(mimeType) {
+    switch (mimeType) {
+        case 'image/png':
+            return 'png';
+        case 'image/jpeg':
+            return 'jpg';
+        case 'image/webp':
+            return 'webp';
+        case 'image/gif':
+            return 'gif';
+        default:
+            return null;
+    }
+}
+function sanitizeArtifactPathSegment(value) {
+    const cleaned = value.replace(/[^a-zA-Z0-9._-]/g, '_').replace(/_+/g, '_');
+    const trimmed = cleaned.replace(/^_+|_+$/g, '');
+    return trimmed.length > 0 ? trimmed.slice(0, 96) : 'item';
+}
+function parseUserImageAttachments(raw) {
+    if (raw === undefined)
+        return [];
+    if (!Array.isArray(raw)) {
+        throw new Error('attachments must be an array when provided');
+    }
+    if (raw.length > USER_IMAGE_ATTACHMENT_MAX_COUNT) {
+        throw new Error(`at most ${String(USER_IMAGE_ATTACHMENT_MAX_COUNT)} image attachments are allowed`);
+    }
+    return raw.map((item, index) => {
+        if (typeof item !== 'object' || item === null || Array.isArray(item)) {
+            throw new Error(`attachments[${String(index)}] must be an object`);
+        }
+        const record = item;
+        const kind = record['kind'];
+        const mimeType = record['mimeType'];
+        const byteLength = record['byteLength'];
+        const dataBase64 = record['dataBase64'];
+        if (kind !== 'image') {
+            throw new Error(`attachments[${String(index)}].kind must be image`);
+        }
+        if (typeof mimeType !== 'string' || userImageMimeTypeToExt(mimeType) === null) {
+            throw new Error(`attachments[${String(index)}].mimeType is unsupported`);
+        }
+        if (typeof byteLength !== 'number' ||
+            !Number.isFinite(byteLength) ||
+            byteLength <= 0 ||
+            byteLength > USER_IMAGE_ATTACHMENT_MAX_BYTES) {
+            throw new Error(`attachments[${String(index)}].byteLength must be between 1 and ${String(USER_IMAGE_ATTACHMENT_MAX_BYTES)}`);
+        }
+        if (typeof dataBase64 !== 'string' || dataBase64.trim() === '') {
+            throw new Error(`attachments[${String(index)}].dataBase64 is required`);
+        }
+        return { kind, mimeType, byteLength, dataBase64 };
+    });
+}
+function isStrictBase64Payload(value) {
+    if (value.length === 0 || value.length % 4 !== 0)
+        return false;
+    return /^(?:[A-Za-z0-9+/]{4})*(?:[A-Za-z0-9+/]{2}==|[A-Za-z0-9+/]{3}=)?$/.test(value);
+}
+function prepareUserImageAttachments(attachments) {
+    return attachments.map((attachment, index) => {
+        if (!isStrictBase64Payload(attachment.dataBase64)) {
+            throw new Error(`attachments[${String(index)}].dataBase64 must be strict base64`);
+        }
+        const bytes = Buffer.from(attachment.dataBase64, 'base64');
+        if (bytes.length !== attachment.byteLength) {
+            throw new Error(`attachments[${String(index)}].byteLength mismatch: declared=${String(attachment.byteLength)} decoded=${String(bytes.length)}`);
+        }
+        return {
+            mimeType: attachment.mimeType,
+            bytes,
+        };
+    });
+}
+async function persistPreparedUserImageAttachments(args) {
+    if (args.attachments.length === 0)
+        return undefined;
+    const eventsBase = persistence_1.DialogPersistence.getDialogEventsPath(args.dialog.id, args.dialog.status);
+    const safeMsgId = sanitizeArtifactPathSegment(args.msgId);
+    const contentItems = [];
+    for (let index = 0; index < args.attachments.length; index += 1) {
+        const attachment = args.attachments[index];
+        const ext = userImageMimeTypeToExt(attachment.mimeType);
+        if (ext === null) {
+            throw new Error(`attachments[${String(index)}].mimeType is unsupported`);
+        }
+        const relPath = path_1.default.posix.join('artifacts', 'user-input', safeMsgId, `${String(index + 1).padStart(2, '0')}-${(0, crypto_1.randomUUID)()}.${ext}`);
+        const absPath = path_1.default.join(eventsBase, ...relPath.split('/'));
+        await promises_1.default.mkdir(path_1.default.dirname(absPath), { recursive: true });
+        await promises_1.default.writeFile(absPath, attachment.bytes);
+        contentItems.push({
+            type: 'input_image',
+            mimeType: attachment.mimeType,
+            byteLength: attachment.bytes.length,
+            artifact: {
+                rootId: args.dialog.id.rootId,
+                selfId: args.dialog.id.selfId,
+                status: args.dialog.status,
+                relPath,
+            },
+        });
+    }
+    return contentItems;
+}
 function parsePersistableDialogStatus(raw) {
     if (raw !== 'running' && raw !== 'completed' && raw !== 'archived') {
         return null;
@@ -230,6 +343,7 @@ async function queueUserSupplementAtGenerationBoundary(dialog, prompt) {
     }
     const queued = dialog.queueUserPromptAtGenerationBoundary({
         prompt: prompt.content,
+        contentItems: prompt.contentItems,
         msgId: prompt.msgId,
         grammar: prompt.grammar,
         userLanguageCode: prompt.userLanguageCode,
@@ -239,6 +353,7 @@ async function queueUserSupplementAtGenerationBoundary(dialog, prompt) {
         course: dialog.currentCourse,
         msgId: queued.msgId,
         content: queued.prompt,
+        contentItems: queued.contentItems,
         grammar: queued.grammar ?? 'markdown',
         origin: 'user',
         userLanguageCode: queued.userLanguageCode,
@@ -1135,6 +1250,7 @@ async function handleDisplayCourse(ws, packet) {
 async function handleUserMsg2Dlg(ws, packet) {
     try {
         const { dialog: dialogIdent, content, msgId } = packet;
+        const attachments = parseUserImageAttachments(packet.attachments);
         const userLanguageCode = resolveUserLanguageCode(ws, packet.userLanguageCode);
         // Basic validation
         if (!dialogIdent || !content || !msgId) {
@@ -1170,6 +1286,7 @@ async function handleUserMsg2Dlg(ws, packet) {
             grammar: 'markdown',
             userLanguageCode,
         };
+        const preparedAttachments = prepareUserImageAttachments(attachments);
         // If the dialog is already active for this WebSocket, runnable (status === 'running'),
         // and has an event forwarder (subChan),
         // drive it directly to preserve in-memory state (pending subdialogs, teammate tellask tracking, etc).
@@ -1187,12 +1304,21 @@ async function handleUserMsg2Dlg(ws, packet) {
             existingDialog.status === 'running' &&
             existingSub &&
             existingSub.dialogKey === existingDialog.id.valueOf()) {
-            const queuedAtBoundary = await queueUserSupplementAtGenerationBoundary(existingDialog, effectivePrompt);
+            const contentItems = await persistPreparedUserImageAttachments({
+                dialog: existingDialog,
+                msgId: effectivePrompt.msgId,
+                attachments: preparedAttachments,
+            });
+            const queuedAtBoundary = await queueUserSupplementAtGenerationBoundary(existingDialog, {
+                ...effectivePrompt,
+                contentItems,
+            });
             if (queuedAtBoundary) {
                 return;
             }
             await (0, kernel_driver_1.driveDialogStream)(existingDialog, {
                 content: effectivePrompt.content,
+                ...(contentItems === undefined ? {} : { contentItems }),
                 msgId: effectivePrompt.msgId,
                 grammar: effectivePrompt.grammar,
                 userLanguageCode: effectivePrompt.userLanguageCode,
@@ -1220,12 +1346,21 @@ async function handleUserMsg2Dlg(ws, packet) {
                 return;
             }
             await setupWebSocketSubscription(ws, dialog);
-            const queuedAtBoundary = await queueUserSupplementAtGenerationBoundary(dialog, effectivePrompt);
+            const contentItems = await persistPreparedUserImageAttachments({
+                dialog,
+                msgId: effectivePrompt.msgId,
+                attachments: preparedAttachments,
+            });
+            const queuedAtBoundary = await queueUserSupplementAtGenerationBoundary(dialog, {
+                ...effectivePrompt,
+                contentItems,
+            });
             if (queuedAtBoundary) {
                 return;
             }
             await (0, kernel_driver_1.driveDialogStream)(dialog, {
                 content: effectivePrompt.content,
+                ...(contentItems === undefined ? {} : { contentItems }),
                 msgId: effectivePrompt.msgId,
                 grammar: effectivePrompt.grammar,
                 userLanguageCode: effectivePrompt.userLanguageCode,
@@ -1401,6 +1536,7 @@ async function handleResumeAll(ws, packet) {
 async function handleReceiveHumanReply(ws, packet) {
     try {
         const { dialog: dialogIdent, content, msgId, questionId, continuationType } = packet;
+        const attachments = parseUserImageAttachments(packet.attachments);
         const userLanguageCode = resolveUserLanguageCode(ws, packet.userLanguageCode);
         // Basic validation
         if (!dialogIdent || !content || !msgId || !questionId) {
@@ -1436,14 +1572,7 @@ async function handleReceiveHumanReply(ws, packet) {
             grammar: 'markdown',
             userLanguageCode,
         };
-        const removed = await persistence_1.DialogPersistence.removeQuestion4HumanState(dialogIdObj, questionId);
-        if (!removed.found) {
-            ws.send(JSON.stringify({
-                type: 'error',
-                message: `Question ${questionId} not found in dialog ${dialogId}`,
-            }));
-            return;
-        }
+        const preparedAttachments = prepareUserImageAttachments(attachments);
         // Restore the canonical dialog instances (root + subdialogs) to avoid duplicates.
         const rootDialog = await (0, dialog_instance_registry_1.getOrRestoreRootDialog)(dialogIdObj.rootId, 'running');
         if (!rootDialog) {
@@ -1460,6 +1589,14 @@ async function handleReceiveHumanReply(ws, packet) {
         }
         // Ensure the requesting WebSocket receives q4h_answered and subsequent resume stream events.
         await setupWebSocketSubscription(ws, dialog);
+        const removed = await persistence_1.DialogPersistence.removeQuestion4HumanState(dialogIdObj, questionId);
+        if (!removed.found) {
+            ws.send(JSON.stringify({
+                type: 'error',
+                message: `Question ${questionId} not found in dialog ${dialogId}`,
+            }));
+            return;
+        }
         const removedQuestion = removed.removedQuestion;
         if (!removedQuestion) {
             throw new Error(`Q4H remove invariant violation: found=true but removedQuestion missing (rootId=${dialog.id.rootId} selfId=${dialog.id.selfId} questionId=${questionId})`);
@@ -1469,6 +1606,11 @@ async function handleReceiveHumanReply(ws, packet) {
             throw new Error(`Q4H remove invariant violation: missing callId on answered question ` +
                 `(rootId=${dialog.id.rootId} selfId=${dialog.id.selfId} questionId=${questionId})`);
         }
+        const contentItems = await persistPreparedUserImageAttachments({
+            dialog,
+            msgId: effectivePrompt.msgId,
+            attachments: preparedAttachments,
+        });
         const askHumanOriginCourse = removedQuestion.callSiteRef.course;
         const askHumanCarryoverContent = (0, inter_dialog_format_1.formatTellaskCarryoverResultContent)({
             originCourse: askHumanOriginCourse,
@@ -1487,6 +1629,7 @@ async function handleReceiveHumanReply(ws, packet) {
             originCourse: (0, storage_1.toCallingCourseNumber)(askHumanOriginCourse),
             calling_genseq: removedQuestion.callSiteRef.callingGenseq,
             carryoverContent: askHumanCarryoverContent,
+            contentItems,
         });
         await dialog.addChatMessages(askHumanResultMirror);
         // Emit q4h_answered event for answered question
@@ -1505,6 +1648,7 @@ async function handleReceiveHumanReply(ws, packet) {
                 prompt: effectivePrompt.content,
                 msgId: effectivePrompt.msgId,
                 grammar: effectivePrompt.grammar,
+                contentItems,
                 userLanguageCode: effectivePrompt.userLanguageCode,
                 q4hAnswerCallId: askHumanCallId,
             });
@@ -1520,6 +1664,7 @@ async function handleReceiveHumanReply(ws, packet) {
         // The continuation input carries correlation only; it does not persist another user prompt fact.
         await (0, kernel_driver_1.driveDialogStream)(dialog, {
             content: effectivePrompt.content,
+            ...(contentItems === undefined ? {} : { contentItems }),
             msgId: effectivePrompt.msgId,
             grammar: effectivePrompt.grammar,
             userLanguageCode: effectivePrompt.userLanguageCode,

package/dist/tools/builtins.js

@@ -16,6 +16,7 @@ const mcp_1 = require("./mcp");
 const mem_1 = require("./mem");
 const os_1 = require("./os");
 const pending_tellask_reminder_1 = require("./pending-tellask-reminder");
+const picture_1 = require("./picture");
 const registry_1 = require("./registry");
 const ripgrep_1 = require("./ripgrep");
 const team_mgmt_1 = require("./team_mgmt");
@@ -50,6 +51,8 @@ function manualSpecFor(toolsetId) {
 (0, registry_1.registerTool)(fs_1.moveFileTool);
 (0, registry_1.registerTool)(fs_1.moveDirTool);
 (0, registry_1.registerTool)(txt_1.readFileTool);
+(0, registry_1.registerTool)(picture_1.readPictureTool);
+(0, registry_1.registerTool)(picture_1.writePictureTool);
 (0, registry_1.registerTool)(txt_1.createNewFileTool);
 (0, registry_1.registerTool)(txt_1.overwriteEntireFileTool);
 (0, registry_1.registerTool)(txt_1.prepareFileRangeEditTool);
@@ -182,6 +185,7 @@ for (const tool of team_mgmt_1.teamMgmtTools) {
 (0, registry_1.registerToolset)('ws_read', [
     fs_1.listDirTool,
     txt_1.readFileTool,
+    picture_1.readPictureTool,
     ripgrep_1.ripgrepFilesTool,
     ripgrep_1.ripgrepSnippetsTool,
     ripgrep_1.ripgrepCountTool,
@@ -191,8 +195,8 @@ for (const tool of team_mgmt_1.teamMgmtTools) {
 (0, registry_1.setToolsetMeta)('ws_read', {
     source: 'dominds',
     descriptionI18n: {
-        en: 'rtws read-only access: list directories, read files, and search code/content to gather facts safely.',
-        zh: '运行时工作区只读访问：列目录、读文件、检索代码与文本，用于安全获取事实。',
+        en: 'rtws read-only access: list directories, read text/images, and search code/content to gather facts safely.',
+        zh: '运行时工作区只读访问：列目录、读文本/图片、检索代码与文本，用于安全获取事实。',
     },
     promptFilesI18n: promptFilesFor('ws_read'),
     manualSpec: manualSpecFor('ws_read'),
@@ -205,6 +209,8 @@ for (const tool of team_mgmt_1.teamMgmtTools) {
     fs_1.moveFileTool,
     fs_1.moveDirTool,
     txt_1.readFileTool,
+    picture_1.readPictureTool,
+    picture_1.writePictureTool,
     txt_1.createNewFileTool,
     txt_1.overwriteEntireFileTool,
     txt_1.prepareFileAppendTool,
@@ -222,8 +228,8 @@ for (const tool of team_mgmt_1.teamMgmtTools) {
 (0, registry_1.setToolsetMeta)('ws_mod', {
     source: 'dominds',
     descriptionI18n: {
-        en: 'rtws read/write access: inspect, create, move, delete, and precisely edit workspace files.',
-        zh: '运行时工作区读写访问：检查、创建、移动、删除，并精确编辑工作区文件。',
+        en: 'rtws read/write access: inspect, create, move, delete, precisely edit text files, and read/write workspace images.',
+        zh: '运行时工作区读写访问：检查、创建、移动、删除、精确编辑文本文件，并读写工作区图片。',
     },
     promptFilesI18n: promptFilesFor('ws_mod'),
     manualSpec: manualSpecFor('ws_mod'),