newmark-agent 0.4.2 → 0.4.4

package/dist/conversation-utility-host.js

@@ -93,6 +93,9 @@ async function handle(request) {
     if (request.method === 'set_mode') {
         return kernel.setMode(checkedTarget(request.params.target), request.params.mode);
     }
+    if (request.method === 'set_model') {
+        return kernel.setModel(checkedTarget(request.params.target), request.params.model);
+    }
     if (request.method === 'set_input_mode') {
         return kernel.setInputMode(checkedTarget(request.params.target), request.params.mode);
     }

package/dist/core/agent.d.ts

@@ -314,6 +314,18 @@ export declare class Agent {
         model: string;
         fallback: boolean;
     } | null;
+    providerUsageTotals: {
+        input: number;
+        output: number;
+        cacheRead: number;
+        cacheWrite: number;
+    };
+    lastProviderUsage: {
+        input: number;
+        output: number;
+        cacheRead: number;
+        cacheWrite: number;
+    };
     private compressionCache;
     private pendingHistoryRemovals;
     private branchMailbox;
@@ -631,12 +643,18 @@ export declare class Agent {
     setConversationPinned(id: string, pinned: boolean): boolean;
     renameConversation(id: string, title: string): boolean;
     /**
-     * 首 Build 命名提示判定：仅当 (1) 当前对话尚无历史 Build（排除当前 run）且
-     * (2) 其持久化 title 仍是自动生成（含为空）时返回 true。满足条件时运行时会
-     * 在首个 provider request 的 bootstrap 注入一次性命名指令，让 Agent 调用
-     * conversation_rename 自行命名。判定本身只读存储、无副作用，保缓存稳定。
+     * 首 Build 命名判定：仅当 (1) 当前对话尚无历史 Build（排除当前 run）且
+     * (2) 其持久化 title 仍是自动生成（含为空）时返回 true。dev-0.4.3 起不再
+     * 用该判定注入首轮 tool-call 指令，而是在首个完成 Build 的最终响应处自动
+     * 命名（见 maybeAutoRenameConversationFromRun）。判定本身只读存储、无副作用。
      */
     shouldPromptConversationRename(): boolean;
+    /**
+     * 从首个 Build 的最终响应中提取一个简短对话标题。跳过 Markdown 标题/列表
+     * 符号与固定 section 标题，取第一条有意义的摘要句并做保守清洗。
+     */
+    private deriveConversationTitleFromSummary;
+    private maybeAutoRenameConversationFromRun;
     reorderConversations(ids: string[]): boolean;
     flushConversationState(): void;
     private titleFromMessages;
@@ -730,6 +748,19 @@ export declare class Agent {
      * 已有的 renameConversation 持久化路径。返回简短结果以保缓存友好。
      */
     handleConversationRename(args: string): NewmarkToolResult;
+    /**
+     * task_read：读取当前对话的持久化内联任务清单（conversationPlan）。
+     * 只读、无副作用，输出有界（每项 text 截断到 240 字符），保缓存友好。
+     * Agent 在需要具体任务状态时调用，替代把动态清单注入每个 provider request。
+     */
+    handleTaskRead(): NewmarkToolResult;
+    /**
+     * task_create：把任务项写入当前对话的持久化内联任务清单（conversationPlan）。
+     * action=create 追加单项；action=update 按 id/index 改状态或文本；action=clear
+     * 移除已完成项。写入走 updateConversationPlan 持久化路径，GUI Task 面板与
+     * TUI plan 视图立即反映。返回紧凑确认，避免回显全量清单。
+     */
+    handleTaskCreate(args: string): NewmarkToolResult;
     private conversationTree;
     private currentRuntimeBranchId;
     handleBranchList(args: string): NewmarkToolResult;
@@ -773,6 +804,12 @@ export declare class Agent {
     modelLabel(): string;
     estimateContextTokens(messages?: Array<Record<string, unknown>>): number;
     private estimateContextTokenComponents;
+    recordProviderUsage(input: {
+        input: number;
+        output: number;
+        cacheRead: number;
+        cacheWrite: number;
+    }): void;
     contextWindow(modelName?: string): {
         estimatedTokens: number;
         maxTokens: number;
@@ -789,6 +826,12 @@ export declare class Agent {
         compressionEnabled?: boolean;
         cacheEntries?: number;
         archiveEntries?: number;
+        providerTotalTokens?: number;
+        providerInputTokens?: number;
+        providerOutputTokens?: number;
+        providerCacheReadTokens?: number;
+        providerCacheWriteTokens?: number;
+        providerCacheReadRatio?: number;
     };
     private resolveWindowModel;
     private contextMaxTokens;
@@ -892,6 +935,11 @@ export declare class Agent {
     modelValidationStatus(): ModelValidationProgress;
     private runModelValidation;
     engineModel(): LLMProvider | null;
+    /**
+     * dev-0.4.3 模型原生思考强度档位映射表（模型名 → thinking_tier_map）。
+     * 未配置映射的模型返回 undefined，provider 侧维持默认透传（不变动映射）。
+     */
+    private modelThinkingTierMaps;
     editorModelRequest(input: {
         path?: string;
         content?: string;

package/dist/core/agent.js

@@ -96,6 +96,11 @@ function throwIfAgentAborted(signal) {
     error.name = 'AbortError';
     throw error;
 }
+/** task_read 输出的单项文本上限：清单项只作状态索引，不承载长描述。 */
+function compactPlanItemText(value) {
+    const clean = String(value || '').replace(/\s+/g, ' ').trim();
+    return clean.length <= 240 ? clean : `${clean.slice(0, 237)}...`;
+}
 let CORE_SYSTEM_PROMPT = `You are Newmark Agent, a powerful AI coding assistant built into a native desktop application.
 
 ## Available Tools
@@ -149,8 +154,8 @@ let CORE_SYSTEM_PROMPT = `You are Newmark Agent, a powerful AI coding assistant
 - When the current instruction is a new task, complete that task without silently appending unrelated historical work.
 
 ## Inline Task Management (Mandatory)
-- For every multi-step conversation task, maintain a compact inline checklist in the current Build work state with actionable items and one status per item: pending, in_progress, completed, or blocked.
-- Update that checklist as work changes and use it to drive tool order and final verification. Keep it bounded to actionable task labels; never expose hidden reasoning.
+- For every multi-step conversation task, persist a compact task checklist through the task_create tool: create actionable items when the work starts, update each item's status (pending|in_progress|done) as work progresses, and mark items done only after verification. These items are the same list the GUI Task panel and TUI plan view render, and they persist across Build Blocks.
+- Call task_read to reload the concrete checklist state whenever you need it; the system prompt intentionally does not inject the live list so the provider prefix cache stays stable. Keep items bounded to actionable task labels; never expose hidden reasoning.
 - The inline checklist is the per-turn task manager. The durable linked-plan document exists and is available through the linked_plan tool when explicitly needed, but its full contents are not injected into every request.
 
 ## Guidelines
@@ -214,6 +219,8 @@ class Agent {
     continuations = [];
     activeConversationId = 'default';
     lastCompression = null;
+    providerUsageTotals = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
+    lastProviderUsage = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
     compressionCache = [];
     pendingHistoryRemovals = [];
     branchMailbox = [];
@@ -2153,8 +2160,10 @@ class Agent {
                 if (run.status !== status)
                     return false;
                 const terminalAt = run.endedAt || endedAt;
-                if (status === 'completed')
+                if (status === 'completed') {
                     this.ensureCompletedWorkRunFinalResult(run);
+                    this.maybeAutoRenameConversationFromRun(run);
+                }
                 const goalAudit = this.auditGoalAtWorkRunEnd(run, status, terminalAt);
                 this.enforceGoalTerminalInvariant(status, goalAudit);
                 this.persistBuildBlockWorkOverview(run, status, terminalAt, goalAudit);
@@ -2202,6 +2211,8 @@ class Agent {
         run.status = status;
         run.endedAt = endedAt;
         run.expanded = true;
+        if (status === 'completed')
+            this.maybeAutoRenameConversationFromRun(run);
         this.activeWorkRunId = '';
         this.finalizingWorkRunId = '';
         this.managedWorkRunIds.delete(run.runId);
@@ -2320,6 +2331,8 @@ class Agent {
             const goalAudit = this.auditGoalAtWorkRunEnd(activeRun, terminalStatus, terminalAt);
             this.enforceGoalTerminalInvariant(terminalStatus, goalAudit);
             this.persistBuildBlockWorkOverview(activeRun, terminalStatus, terminalAt, goalAudit);
+            if (terminalStatus === 'completed')
+                this.maybeAutoRenameConversationFromRun(activeRun);
         }
         return event;
     }
@@ -3414,10 +3427,10 @@ class Agent {
         return true;
     }
     /**
-     * 首 Build 命名提示判定：仅当 (1) 当前对话尚无历史 Build（排除当前 run）且
-     * (2) 其持久化 title 仍是自动生成（含为空）时返回 true。满足条件时运行时会
-     * 在首个 provider request 的 bootstrap 注入一次性命名指令，让 Agent 调用
-     * conversation_rename 自行命名。判定本身只读存储、无副作用，保缓存稳定。
+     * 首 Build 命名判定：仅当 (1) 当前对话尚无历史 Build（排除当前 run）且
+     * (2) 其持久化 title 仍是自动生成（含为空）时返回 true。dev-0.4.3 起不再
+     * 用该判定注入首轮 tool-call 指令，而是在首个完成 Build 的最终响应处自动
+     * 命名（见 maybeAutoRenameConversationFromRun）。判定本身只读存储、无副作用。
      */
     shouldPromptConversationRename() {
         if (this.conversationBuildHistory(1).length > 0)
@@ -3431,6 +3444,46 @@ class Agent {
         const messages = entry?.chatMessages || this.chatMessages;
         return this.isGeneratedConversationTitle(priorTitle, conversationId, messages);
     }
+    /**
+     * 从首个 Build 的最终响应中提取一个简短对话标题。跳过 Markdown 标题/列表
+     * 符号与固定 section 标题，取第一条有意义的摘要句并做保守清洗。
+     */
+    deriveConversationTitleFromSummary(summary) {
+        const clean = this.sanitizeAssistantOutput(summary || '').replace(/\r/g, '');
+        const lines = clean.split('\n').map(line => line.trim()).filter(Boolean);
+        for (const line of lines) {
+            const withoutHeading = line
+                .replace(/^#{1,6}\s*/, '')
+                .replace(/^[-*+>]\s*/, '')
+                .trim();
+            if (!withoutHeading)
+                continue;
+            if (/^(做了什么|验证|文件|问题|下一步|What changed|Verification|Files|Issues|Next)[:：]?$/i.test(withoutHeading))
+                continue;
+            const firstSentence = withoutHeading.split(/[。！？!?.;；]/)[0].trim() || withoutHeading;
+            const title = firstSentence
+                .replace(/[{}[\]()<>"'`]/g, '')
+                .replace(/\s+/g, ' ')
+                .trim()
+                .slice(0, 48);
+            if (title.length >= 2)
+                return title;
+        }
+        return '';
+    }
+    maybeAutoRenameConversationFromRun(run) {
+        if (run.status !== 'completed')
+            return;
+        if (!this.shouldPromptConversationRename())
+            return;
+        const finalEvent = [...run.events].reverse().find(event => event.type === 'final_response');
+        const finalMessage = [...this.chatMessages].reverse().find(message => message.role === 'assistant' && message.runId === run.runId);
+        const raw = finalEvent?.content || finalMessage?.content || '';
+        const summary = this.sanitizePublicWorkContent(raw).slice(0, 2000);
+        const title = this.deriveConversationTitleFromSummary(summary);
+        if (title)
+            this.renameConversation(this.activeConversationId || 'default', title);
+    }
     reorderConversations(ids) {
         const prefix = this.workspaceConversationPrefix() || '';
         const normalized = Array.from(new Set((Array.isArray(ids) ? ids : []).map(id => this.safeConversationId(id)).filter(Boolean)));
@@ -4124,8 +4177,8 @@ class Agent {
         const tool = String(input.tool || '').trim();
         if (!tool)
             return { ok: false, output: '[background_tool] tool is required.', error: 'tool is required.' };
-        if (tool === 'background_tool' || tool === 'read_tool_result' || tool === 'compress_tool_result' || tool === 'goal_manage' || tool === 'conversation_rename') {
-            return { ok: false, output: '[background_tool] cannot background the control tools (background_tool/read_tool_result/compress_tool_result/goal_manage/conversation_rename).', error: 'control-tool-unsupported.' };
+        if (tool === 'background_tool' || tool === 'read_tool_result' || tool === 'compress_tool_result' || tool === 'goal_manage' || tool === 'conversation_rename' || tool === 'task_read' || tool === 'task_create') {
+            return { ok: false, output: '[background_tool] cannot background the control tools (background_tool/read_tool_result/compress_tool_result/goal_manage/conversation_rename/task_read/task_create).', error: 'control-tool-unsupported.' };
         }
         // 禁止后台化子代理/编排类工具——它们有独立的生命周期管理，后台化会破坏其
         // mailbox/abort/persistence 语义。
@@ -4280,6 +4333,106 @@ class Agent {
             metadata: { kind: 'conversation-rename' },
         };
     }
+    /**
+     * task_read：读取当前对话的持久化内联任务清单（conversationPlan）。
+     * 只读、无副作用，输出有界（每项 text 截断到 240 字符），保缓存友好。
+     * Agent 在需要具体任务状态时调用，替代把动态清单注入每个 provider request。
+     */
+    handleTaskRead() {
+        const plan = this.normalizeConversationPlan(this.conversationPlan);
+        const items = plan.items.map((item, index) => ({
+            index,
+            id: item.id,
+            status: item.status,
+            task: compactPlanItemText(item.text),
+            updatedAt: item.updatedAt || '',
+        }));
+        return {
+            ok: true,
+            output: JSON.stringify({
+                ok: true,
+                conversationId: this.activeConversationId || 'default',
+                total: items.length,
+                unfinished: items.filter(item => item.status !== 'done').length,
+                items,
+            }, null, 2),
+            metadata: { kind: 'task-read' },
+        };
+    }
+    /**
+     * task_create：把任务项写入当前对话的持久化内联任务清单（conversationPlan）。
+     * action=create 追加单项；action=update 按 id/index 改状态或文本；action=clear
+     * 移除已完成项。写入走 updateConversationPlan 持久化路径，GUI Task 面板与
+     * TUI plan 视图立即反映。返回紧凑确认，避免回显全量清单。
+     */
+    handleTaskCreate(args) {
+        let input = {};
+        try {
+            input = JSON.parse(args || '{}');
+        }
+        catch { }
+        const action = String(input.action || 'create').trim();
+        const plan = this.normalizeConversationPlan(this.conversationPlan);
+        const now = new Date().toISOString();
+        if (action === 'create') {
+            const text = String(input.task || input.text || '').replace(/\s+/g, ' ').trim();
+            if (!text)
+                return { ok: false, output: '[task_create] task text is required.', error: 'task text is required.' };
+            const item = {
+                id: `plan-${Date.now()}-${Math.random().toString(16).slice(2, 6)}`,
+                text: text.slice(0, 400),
+                status: 'pending',
+                createdAt: now,
+                updatedAt: now,
+            };
+            plan.items.push(item);
+            this.updateConversationPlan(plan);
+            return {
+                ok: true,
+                output: JSON.stringify({ ok: true, action, id: item.id, status: item.status, total: plan.items.length }, null, 2),
+                metadata: { kind: 'task-create' },
+            };
+        }
+        if (action === 'update') {
+            const id = String(input.id || '').trim();
+            const index = Number(input.index);
+            const status = String(input.status || '').trim();
+            const text = String(input.task || input.text || '').replace(/\s+/g, ' ').trim();
+            const target = id
+                ? plan.items.find(item => item.id === id)
+                : Number.isInteger(index) && index >= 0 && index < plan.items.length
+                    ? plan.items[index]
+                    : undefined;
+            if (!target)
+                return { ok: false, output: '[task_create] no matching task item for update (pass id or valid index).', error: 'task item not found.' };
+            if (status) {
+                if (!['pending', 'in_progress', 'done', 'blocked'].includes(status)) {
+                    return { ok: false, output: '[task_create] status must be pending|in_progress|done|blocked.', error: 'invalid status.' };
+                }
+                target.status = status === 'blocked' ? 'pending' : status;
+            }
+            if (text)
+                target.text = text.slice(0, 400);
+            target.updatedAt = now;
+            this.updateConversationPlan(plan);
+            return {
+                ok: true,
+                output: JSON.stringify({ ok: true, action, id: target.id, status: target.status, total: plan.items.length }, null, 2),
+                metadata: { kind: 'task-create' },
+            };
+        }
+        if (action === 'clear') {
+            const remaining = plan.items.filter(item => item.status !== 'done');
+            const removed = plan.items.length - remaining.length;
+            this.updateConversationPlan({ items: remaining });
+            return {
+                ok: true,
+                output: JSON.stringify({ ok: true, action, removed, total: remaining.length }, null, 2),
+                metadata: { kind: 'task-create' },
+            };
+        }
+        return { ok: false, output: '[task_create] action must be create|update|clear.', error: 'invalid action.' };
+    }
     conversationTree() {
         const stateKey = this.workspaceConversationStateKey();
         const stored = this.readStoredConversationState();
@@ -5099,6 +5252,20 @@ class Agent {
             buildBlockTokens: estimate(buildBlockAsciiChars, buildBlockNonAsciiChars, buildBlockStructuralChars, true),
         };
     }
+    recordProviderUsage(input) {
+        const bounded = (value) => Math.max(0, Math.floor(Number(value) || 0));
+        const usage = {
+            input: bounded(input.input),
+            output: bounded(input.output),
+            cacheRead: bounded(input.cacheRead),
+            cacheWrite: bounded(input.cacheWrite),
+        };
+        this.lastProviderUsage = usage;
+        this.providerUsageTotals.input += usage.input;
+        this.providerUsageTotals.output += usage.output;
+        this.providerUsageTotals.cacheRead += usage.cacheRead;
+        this.providerUsageTotals.cacheWrite += usage.cacheWrite;
+    }
     contextWindow(modelName = this.model) {
         const estimatedTokens = this.estimateContextTokens();
         // Display and compression must share one window resolution. Both resolve
@@ -5127,12 +5294,32 @@ class Agent {
             compressionEnabled: this.config.getBool('context', 'auto_compress'),
             cacheEntries: this.compressionCache.length,
             archiveEntries: this.compressionArchiveEntryCount(),
+            providerTotalTokens: this.providerUsageTotals.input + this.providerUsageTotals.output,
+            providerInputTokens: this.providerUsageTotals.input,
+            providerOutputTokens: this.providerUsageTotals.output,
+            providerCacheReadTokens: this.providerUsageTotals.cacheRead,
+            providerCacheWriteTokens: this.providerUsageTotals.cacheWrite,
+            providerCacheReadRatio: this.providerUsageTotals.input > 0
+                ? Math.min(1, this.providerUsageTotals.cacheRead / this.providerUsageTotals.input)
+                : 0,
         };
     }
     resolveWindowModel(modelName) {
-        if (modelName !== 'auto')
-            return this.config.findModel(modelName);
-        return this.activeModelConfig() || this.config.findModel(this.config.getStr('models', 'default_model'));
+        // The context window (display ring, inspector, and compaction trigger) must
+        // resolve the deployment that is actually running. For the active selection
+        // — auto or a fixed model — resolve through the active deployment so a
+        // qualified selection or two same-named models across providers never fall
+        // through to the 128000 default. Only a caller-supplied foreign name (for
+        // example a validation probe against another model) resolves by bare name.
+        if (modelName === 'auto' || modelName === this.model || modelName === this.activeModelName()) {
+            const active = this.activeModelConfig();
+            if (active)
+                return active;
+        }
+        const byName = this.config.findModel(modelName);
+        if (byName)
+            return byName;
+        return this.config.findModel(this.config.getStr('models', 'default_model'));
     }
     contextMaxTokens(modelName = this.model) {
         const model = this.resolveWindowModel(modelName);
@@ -6368,7 +6555,7 @@ class Agent {
             this.modelValidationProgress = { ...this.modelValidationProgress, currentModel, currentCheck: 'catalog' };
             const inferredVision = !!m.vision || (0, config_1.inferModelVisionCapability)(m.name, m.display, m.description, m.provider, m.provider_protocol);
             const inferredImageOutput = !!m.image_output || /(?:^|[-_.])(gpt-image|dall-e|imagen|imagegen|image-generation)(?:$|[-_.])/i.test(m.name);
-            const p = new provider_1.LLMProvider(m.provider, m.provider_url, m.api_key, m.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'));
+            const p = new provider_1.LLMProvider(m.provider, m.provider_url, m.api_key, m.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'), undefined, this.modelThinkingTierMaps(m));
             let catalog = catalogByProvider.get(m.provider_id);
             if (!catalog && m.provider_url && m.api_key) {
                 try {
@@ -6484,7 +6671,17 @@ class Agent {
         const m = this.activeModelConfig();
         if (!m)
             return null;
-        return new provider_1.LLMProvider(m.provider, m.provider_url, m.api_key, m.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'));
+        return new provider_1.LLMProvider(m.provider, m.provider_url, m.api_key, m.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'), undefined, this.modelThinkingTierMaps(m));
+    }
+    /**
+     * dev-0.4.3 模型原生思考强度档位映射表（模型名 → thinking_tier_map）。
+     * 未配置映射的模型返回 undefined，provider 侧维持默认透传（不变动映射）。
+     */
+    modelThinkingTierMaps(model) {
+        const map = model?.thinking_tier_map;
+        if (!model?.name || !map || typeof map !== 'object' || !Object.keys(map).length)
+            return undefined;
+        return { [model.name]: map };
     }
     async editorModelRequest(input, signal) {
         const models = this.config.allModels().filter(model => {
@@ -6510,8 +6707,8 @@ class Agent {
         if (!selected?.api_key || !selected.provider_url)
             return { ok: false, text: '', error: 'No available editor prediction model.' };
         const provider = input.completion
-            ? new provider_1.LLMProvider(selected.provider, selected.provider_url, selected.api_key, selected.provider_protocol, 'chat_stream', this.config.contextFlag('provider_adapters_v2'), EDITOR_COMPLETION_TIMEOUT_MS)
-            : new provider_1.LLMProvider(selected.provider, selected.provider_url, selected.api_key, selected.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'));
+            ? new provider_1.LLMProvider(selected.provider, selected.provider_url, selected.api_key, selected.provider_protocol, 'chat_stream', this.config.contextFlag('provider_adapters_v2'), EDITOR_COMPLETION_TIMEOUT_MS, this.modelThinkingTierMaps(selected))
+            : new provider_1.LLMProvider(selected.provider, selected.provider_url, selected.api_key, selected.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'), undefined, this.modelThinkingTierMaps(selected));
         const language = path.extname(String(input.path || '')).replace(/^\./, '') || 'text';
         const system = input.completion
             ? 'You are an inline code completion engine. Return only the exact text to insert at the cursor. Do not use Markdown fences or explanations.'
@@ -6807,20 +7004,70 @@ class Agent {
                     throw new Error(message);
                 }
             }
-            const text = typeof input === 'string' ? input : String(input.text || '');
+            let text = typeof input === 'string' ? input : String(input.text || '');
             const inputEnvelope = typeof input === 'string' ? null : input;
-            const hiddenUserInput = inputEnvelope?.hiddenUserInput === true;
-            // An empty selection may be repaired to the configured default, but an
-            // explicitly requested fixed model must remain visible to the
-            // fail-closed availability check below. Otherwise a typo such as
-            // provider-that-does-not-exist/missing-model silently becomes the first
-            // configured fixture/default deployment.
+            let hiddenUserInput = inputEnvelope?.hiddenUserInput === true;
             const explicitFixedModel = this.model !== '' && this.model !== 'auto';
             if (!explicitFixedModel)
                 this.ensureUsableModelSelection();
-            const clientMessageId = String(inputEnvelope?.clientMessageId || '').trim();
+            let clientMessageId = String(inputEnvelope?.clientMessageId || '').trim();
             const inputRunId = String(inputEnvelope?.runId || this.activeWorkRunId || '').trim();
-            const rawImages = typeof input === 'string' ? [] : (Array.isArray(input.images) ? input.images : []);
+            let rawImages = typeof input === 'string' ? [] : (Array.isArray(input.images) ? input.images : []);
+            // dev-0.4.3 Guide 批量续接：同一 Build block 内连续到达的多个 Guide
+            // 由 conversation kernel 合并到这里一次注入，不再每来一个 Guide 响应一次。
+            const batchGuides = Array.isArray(inputEnvelope?.batchGuides) ? inputEnvelope.batchGuides : [];
+            if (batchGuides.length) {
+                const batchRunId = inputRunId || this.activeWorkRunId || '';
+                const batchTarget = this.currentConversationTarget();
+                const appliedAt = this.nowIso();
+                const persisted = [];
+                const batchImages = [];
+                for (const guide of batchGuides) {
+                    const guideClientMessageId = String(guide.clientMessageId || '').trim();
+                    if (!guideClientMessageId)
+                        continue;
+                    let guideImages = [];
+                    let guideAttachments = [];
+                    try {
+                        const prepared = this.prepareSubmittedConversationImages(guide.images);
+                        guideImages = prepared.images;
+                        guideAttachments = prepared.attachments;
+                    }
+                    catch (error) {
+                        this.status = 'idle';
+                        return [{ type: 'text', text: `[Attachment rejected] ${error instanceof Error ? error.message : String(error)}` }];
+                    }
+                    const guideDisplay = guideImages.length
+                        ? `${guide.text}${guide.text ? '\n\n' : ''}[${guideImages.length} image attachment${guideImages.length === 1 ? '' : 's'}]`
+                        : guide.text;
+                    batchImages.push(...guideImages);
+                    // 图片统一由主 hidden user 消息携带，Guide 自身 history 只保留文本，
+                    // 避免同一批图片在上下文中重复计费。
+                    this.persistGuideMessage(guideClientMessageId, guideDisplay, batchRunId, guide.text, guideAttachments, String(guide.guideId || ''));
+                    this.recordGuideReceipt({
+                        clientMessageId: guideClientMessageId,
+                        guideId: String(guide.guideId || '') || undefined,
+                        target: batchTarget,
+                        runId: batchRunId,
+                        status: 'applied',
+                        content: guideDisplay,
+                        createdAt: appliedAt,
+                        updatedAt: appliedAt,
+                        appliedAt,
+                    });
+                    this.consumeConversationContinuation({ content: guide.text, queueMode: 'steer', clientMessageId: guideClientMessageId });
+                    persisted.push({ text: guide.text, clientMessageId: guideClientMessageId });
+                }
+                if (persisted.length === 1) {
+                    text = persisted[0].text;
+                }
+                else {
+                    text = `Apply the following intervening Guides in submission order within the current Build Block and continue automatically:\n${persisted.map((guide, index) => `Guide ${index + 1}: ${guide.text}`).join('\n')}`;
+                }
+                hiddenUserInput = true;
+                clientMessageId = '';
+                rawImages = batchImages;
+            }
             let autoRouteEvaluated = false;
             let attachments = [];
             let images = [];
@@ -7639,7 +7886,7 @@ class Agent {
         const activeModel = this.activeModelConfig();
         const activeProvider = this.engineModel();
         const assignedProvider = assignedModel && assignedModel.provider_id !== activeModel?.provider_id
-            ? new provider_1.LLMProvider(assignedModel.provider, assignedModel.provider_url, assignedModel.api_key, assignedModel.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'))
+            ? new provider_1.LLMProvider(assignedModel.provider, assignedModel.provider_url, assignedModel.api_key, assignedModel.provider_protocol, this.config.openAIApiMode(), this.config.contextFlag('provider_adapters_v2'), undefined, this.modelThinkingTierMaps(assignedModel))
             : activeProvider;
         if (!assignedProvider || !model) {
             throw new Error('No LLM configured. Add provider in Settings > Models.');
@@ -8623,7 +8870,7 @@ class Agent {
             '- Memory Lab is governed by an explicit Policy chain: pre-think whether memory is needed; prefer bounded memory_lab_query retrieval; then choose ADD/UPDATE/DELETE only when the user authorizes durable memory mutation.',
             '- Before memory_lab_update, inspect the target with memory_lab_query or memory_lab_read. For an existing component pass expectedUpdatedAt so concurrent/stale writes fail closed; preserve established tag parent paths.',
             '- Use memory_lab_delete only for an explicit user request to forget/remove memory. Prior revisions are retained under Memory Lab/archive and mutation decisions are appended to policy.jsonl for replay.',
-            '- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. Use build_history_query only when the current user asks what specifically happened in one Build Block; querying history is read-only and never authorizes resuming that work.',
+            '- Build history disclosure is two-layered. The request prompt contains only each historical Build Block user input, final summary, and completion status. When the current task continues, fixes, verifies, or depends on earlier Build Blocks, proactively call build_history_query to read the concrete tool activity and results of the relevant block, and reuse that information instead of re-investigating (re-running commands or re-reading files) from scratch. Querying history is read-only and never authorizes resuming that work; do not query merely to answer completion status already shown in the prompt.',
             '- Linked plan disclosure: a durable conversation-linked Markdown plan exists and can be inspected or updated with linked_plan when explicitly needed or required by Plan mode. Its full Markdown and revision are not injected into every model request.',
             '- A memory_lab_update, memory_lab_delete, or memory_lab_reindex call is unfinished until its awaited tool result contains rebuildReceipt.completed=true. The completion receipt is represented by the tool activity inside the current Build block and should not be repeated as a separate completion message.',
             `- Skills and subagents: skill searches enabled metadata and loads one SKILL.md body on demand; skill_download installs offline skill folders; task creates constrained subagents tracked in agent state.`,