evolclaw 3.1.4 → 3.1.6

package/dist/agents/claude-runner.js

@@ -10,6 +10,114 @@ import os from 'os';
 import { logger } from '../utils/logger.js';
 import { checkBlacklist, checkReadonly, summarizeToolInput } from '../core/permission.js';
 import { encodePath } from '../utils/cross-platform.js';
+// ── 模型别名解析 ──
+// SDK 内置的别名表可能落后于代理实际可用的最新模型，
+// 因此优先从 {baseUrl}/models 动态获取各系列最新版本，失败则回退静态表。
+// 已验证可用但尚未出现在 /models 列表中的模型 ID 会被注入候选列表，
+// 等列表更新后注入自动变为 no-op。
+const MODEL_FAMILIES = ['opus', 'sonnet', 'haiku'];
+/** 已验证可用但可能尚未出现在 /models 列表中的模型 ID（注入候选） */
+const INJECTED_MODELS = [];
+/** 静态回退表：动态获取失败时使用 */
+const STATIC_MODEL_ALIASES = {
+    'opus': 'claude-opus-4-8',
+    'sonnet': 'claude-sonnet-4-6',
+    'haiku': 'claude-haiku-4-5-20251001',
+};
+const MODEL_ALIAS_TTL_MS = 5 * 60 * 1000; // 5min
+const modelAliasCache = new Map(); // key: baseUrl
+const modelAliasInFlight = new Set(); // 去重并发刷新
+/** 从模型 ID 列表中提取各 claude 系列的最新版本（按 major.minor 取最高） */
+function deriveAliasesFromModelIds(ids) {
+    // 注入已验证可用的模型（如果列表中已有则去重无影响）
+    const allIds = [...new Set([...ids, ...INJECTED_MODELS])];
+    const best = {};
+    for (const id of allIds) {
+        const m = id.match(/^claude-(opus|sonnet|haiku)-(\d+)-(\d+)/);
+        if (!m)
+            continue;
+        const [, family, majorStr, minorStr] = m;
+        const major = parseInt(majorStr, 10);
+        const minor = parseInt(minorStr, 10);
+        const cur = best[family];
+        if (!cur || major > cur.major || (major === cur.major && minor > cur.minor)) {
+            best[family] = { id, major, minor };
+        }
+    }
+    const aliases = {};
+    for (const [family, info] of Object.entries(best))
+        aliases[family] = info.id;
+    return aliases;
+}
+/** 异步刷新某 baseUrl 的别名缓存（失败静默，不抛出） */
+async function refreshModelAliases(baseUrl, apiKey) {
+    if (modelAliasInFlight.has(baseUrl))
+        return;
+    modelAliasInFlight.add(baseUrl);
+    try {
+        const url = `${baseUrl.replace(/\/$/, '')}/models`;
+        const controller = new AbortController();
+        const timer = setTimeout(() => controller.abort(), 5000);
+        const resp = await fetch(url, {
+            signal: controller.signal,
+            headers: apiKey ? { Authorization: `Bearer ${apiKey}` } : {},
+        });
+        clearTimeout(timer);
+        if (!resp.ok)
+            return;
+        const json = await resp.json();
+        const ids = Array.isArray(json?.data)
+            ? json.data.map((m) => m?.id).filter((x) => typeof x === 'string')
+            : [];
+        const aliases = deriveAliasesFromModelIds(ids);
+        if (ids.length > 0 || Object.keys(aliases).length > 0) {
+            modelAliasCache.set(baseUrl, { aliases, ids, fetchedAt: Date.now() });
+            logger.info(`[AgentRunner] Refreshed models from ${url}: ${ids.length} ids, aliases ${JSON.stringify(aliases)}`);
+        }
+    }
+    catch {
+        // 网络/解析失败：保持静态回退，不打断查询
+    }
+    finally {
+        modelAliasInFlight.delete(baseUrl);
+    }
+}
+/** 将短别名展开为完整 model ID，已是完整 ID 则原样返回 */
+function resolveModelAlias(model, baseUrl) {
+    // 非短别名（已经是完整 ID）直接返回
+    if (!MODEL_FAMILIES.includes(model))
+        return model;
+    // 优先使用动态缓存
+    if (baseUrl) {
+        const cached = modelAliasCache.get(baseUrl);
+        if (cached && (Date.now() - cached.fetchedAt < MODEL_ALIAS_TTL_MS)) {
+            return cached.aliases[model] || STATIC_MODEL_ALIASES[model] || model;
+        }
+    }
+    // 回退静态表
+    return STATIC_MODEL_ALIASES[model] || model;
+}
+/** 支持 1M 上下文窗口的模型 ID 前缀（SDK 通过 `[1m]` 后缀启用）。 */
+const ONE_M_CONTEXT_PREFIXES = ['claude-opus-4-8', 'claude-sonnet-4-6'];
+/**
+ * 为支持 1M 上下文的模型追加 `[1m]` 后缀——仅在交给 SDK query() 时调用。
+ * 目录与校验层始终使用不带后缀的基础 ID，避免与网关 /models 返回值（无 `[1m]`）冲突。
+ */
+function applyContextWindow(modelId) {
+    if (/\[1m\]$/.test(modelId))
+        return modelId; // 已带后缀
+    if (ONE_M_CONTEXT_PREFIXES.some(p => modelId === p))
+        return `${modelId}[1m]`;
+    return modelId;
+}
+/** 根据 SDK model 串（含 [1m] 后缀）返回合适的 autoCompactWindow 值。 */
+function contextWindowFor(sdkModel) {
+    return /\[1m\]$/.test(sdkModel) ? 900000 : 200000;
+}
+/** 解析别名 + 追加 1M 后缀，得到最终交给 SDK 的 model 串。 */
+function resolveSdkModel(model, baseUrl) {
+    return applyContextWindow(resolveModelAlias(model, baseUrl));
+}
 class MessageStream {
     queue = [];
     waiting = null;
@@ -89,6 +197,9 @@ export class AgentRunner {
     permissionContexts = new Map();
     currentEvolclawSessionId;
     claudeExecutablePath;
+    /** 每个 session 最近的子进程 stderr 行（环形缓冲），用于子进程崩溃时还原真正原因 */
+    recentStderr = new Map();
+    static STDERR_BUFFER_MAX = 80;
     constructor(apiKey, model, onSessionIdUpdate, baseUrl, config) {
         this.apiKey = apiKey;
         this.model = model || 'sonnet';
@@ -102,11 +213,17 @@ export class AgentRunner {
         }
     }
     getAgentEnv() {
+        // SDK 0.3.x 起，CLI 在以 root 运行时会拒绝 --dangerously-skip-permissions
+        // （bypassPermissions 模式映射而来），报错 "cannot be used with root/sudo privileges"
+        // 并以 code 1 退出。IS_SANDBOX=1 是 CLI 提供的 root 守卫豁免开关。
+        // 仅在以 root 运行时注入，非 root 部署行为不变。
+        const isRoot = typeof process.getuid === 'function' && process.getuid() === 0;
         return {
             ...process.env,
             ANTHROPIC_AUTH_TOKEN: this.apiKey,
             PATH: process.env.PATH,
             DISABLE_AUTOUPDATER: '1',
+            ...(isRoot ? { IS_SANDBOX: '1' } : {}),
             ...(this.baseUrl ? { ANTHROPIC_BASE_URL: this.baseUrl } : {}),
             ...(this.currentEvolclawSessionId ? { EVOLCLAW_SESSION_ID: this.currentEvolclawSessionId } : {}),
         };
@@ -117,8 +234,28 @@ export class AgentRunner {
     getModel() {
         return this.model;
     }
-    listModels() {
-        return ['opus', 'sonnet', 'haiku'];
+    async listModels() {
+        if (this.baseUrl) {
+            let cached = modelAliasCache.get(this.baseUrl);
+            const stale = !cached || (Date.now() - cached.fetchedAt > MODEL_ALIAS_TTL_MS);
+            // 缓存为空（首次打开）→ 等待刷新；缓存仅过期 → 后台刷新不阻塞
+            if (!cached) {
+                await refreshModelAliases(this.baseUrl, this.apiKey);
+                cached = modelAliasCache.get(this.baseUrl);
+            }
+            else if (stale) {
+                refreshModelAliases(this.baseUrl, this.apiKey);
+            }
+            // 有缓存时返回网关 /models 的全量原始 ID
+            if (cached && cached.ids.length > 0)
+                return cached.ids;
+        }
+        // 无 baseUrl / 刷新超时或失败 → 回退短别名
+        return Object.values(STATIC_MODEL_ALIASES);
+    }
+    /** 将短别名解析为当前代理实际使用的完整 model ID（仅用于展示，不改变持久化值） */
+    resolveModelId(model) {
+        return resolveModelAlias(model, this.baseUrl);
     }
     setEffort(effort) {
         this.effort = effort;
@@ -157,7 +294,7 @@ export class AgentRunner {
     toSdkPermissionMode() {
         const map = {
             'auto': 'auto', // AI 分类器自动判断
-            'bypass': 'default', // 全部自动放行（通过 canUseTool 一律 allow，保留 hook 安全检查）
+            'bypass': 'bypassPermissions', // 全部自动放行（SDK 跳过分类器，canUseTool 仍保留 hook 安全检查）
             'request': 'default', // 部分自动，部分询问
             'edit': 'acceptEdits',
             'plan': 'plan',
@@ -213,49 +350,70 @@ export class AgentRunner {
         if (!adapterHasInteractionPath) {
             return this.handleAskUserQuestionFallback(sessionId, input, questions);
         }
+        // 立即暂停 idle 监控，不等卡片发完再 register
+        permCtx.interactionRouter?.markWaiting(sessionId);
+        let waitMarked = true;
         const answers = {};
-        // 从 permCtx 构造 per-session 的发送函数，避免全局 sendPromptFn 被其他 channel 实例覆盖
-        // 注意：sendPromptFn 是全局单例，多 channel 并发时会被覆盖，导致提示发到错误 channel
         const sendPrompt = permCtx.adapter && permCtx.channelId
             ? async (text) => permCtx.adapter.send(buildEnvelope({ channel: permCtx.adapter.channelName, channelId: permCtx.channelId, replyContext: permCtx.replyContext }), { kind: 'result.text', text, isFinal: true })
             : this.sendPromptFn;
-        // 逐个 question 发送卡片并等待用户选择
         for (let i = 0; i < questions.length; i++) {
             const q = questions[i];
             const requestId = `ask-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
             const cardTitle = q.header ? `💬 ${q.header}` : `💬 问题 ${i + 1}/${questions.length}`;
-            // 统一使用 action 按钮卡片（单选 / 多选均用按钮）
-            const bodyLines = [q.question];
-            if (q.options.some(opt => opt.description)) {
-                bodyLines.push('');
-                q.options.forEach((opt, idx) => {
-                    bodyLines.push(`${idx + 1}. **${opt.label}**${opt.description ? ` — ${opt.description}` : ''}`);
-                });
+            let interaction;
+            if (q.multiSelect) {
+                // 多选：使用 checkers + form 提交（JSON 2.0 CardKit 路径）
+                interaction = {
+                    type: 'interaction',
+                    id: requestId,
+                    kind: {
+                        kind: 'action',
+                        title: cardTitle,
+                        body: q.question,
+                        checkers: q.options.map(opt => ({
+                            key: opt.label,
+                            label: opt.label,
+                            description: opt.description,
+                        })),
+                        buttons: [
+                            { key: 'submit', label: '✅ 确认选择', style: 'primary' },
+                        ],
+                        allowCustomInput: true,
+                    },
+                    channelId: permCtx.channelId,
+                    sessionId,
+                    expiresAt: Date.now() + 5 * 60 * 1000,
+                };
             }
-            const interaction = {
-                type: 'interaction',
-                id: requestId,
-                kind: {
-                    kind: 'action',
-                    title: cardTitle,
-                    body: bodyLines.join('\n'),
-                    buttons: [
-                        ...q.options.map(opt => ({
+            else {
+                // 单选：保持按钮模式
+                const bodyLines = [q.question];
+                if (q.options.some(opt => opt.description)) {
+                    bodyLines.push('');
+                    q.options.forEach((opt, idx) => {
+                        bodyLines.push(`${idx + 1}. **${opt.label}**${opt.description ? ` — ${opt.description}` : ''}`);
+                    });
+                }
+                interaction = {
+                    type: 'interaction',
+                    id: requestId,
+                    kind: {
+                        kind: 'action',
+                        title: cardTitle,
+                        body: bodyLines.join('\n'),
+                        buttons: q.options.map(opt => ({
                             key: opt.label,
                             label: opt.label,
                             style: 'default',
                         })),
-                        ...(permCtx.interceptNextMessage ? [{
-                                key: '_custom_input',
-                                label: '✏️ 手动输入',
-                                style: 'default',
-                            }] : []),
-                    ],
-                },
-                channelId: permCtx.channelId,
-                sessionId,
-                expiresAt: Date.now() + 5 * 60 * 1000,
-            };
+                        allowCustomInput: true,
+                    },
+                    channelId: permCtx.channelId,
+                    sessionId,
+                    expiresAt: Date.now() + 5 * 60 * 1000,
+                };
+            }
             let cardSent = false;
             try {
                 const envelope = buildEnvelope({
@@ -275,7 +433,6 @@ export class AgentRunner {
                 logger.warn(`[AgentRunner] AskUserQuestion card send failed for q${i}:`, err);
             }
             if (!cardSent) {
-                // 卡片发送失败，以纯文本展示选项并自动选推荐项
                 const firstLabel = q.options[0]?.label || '';
                 answers[q.question] = q.multiSelect ? [firstLabel] : firstLabel;
                 if (sendPrompt) {
@@ -284,35 +441,39 @@ export class AgentRunner {
                 }
                 continue;
             }
-            // 等待用户交互
+            // 等待用户交互：先 register 接管计数，再 unmark 占位，消除空窗期
+            // （unmark 必须在 register 之后，否则计数短暂降为 0 触发 onWaitEnd→resume，idle 时钟被重置）
             const answer = await new Promise((resolve) => {
                 permCtx?.interactionRouter?.register(requestId, sessionId, (action, values) => {
                     if (action === 'cancel') {
                         resolve(null);
                     }
-                    else if (action === '_custom_input' && permCtx.interceptNextMessage) {
-                        // "手动输入"：发提示，拦截下一条消息
-                        const sendHint = async () => {
-                            if (sendPrompt) {
-                                await sendPrompt('✏️ 请输入你的想法，回复后继续……');
+                    else if (action === '_custom_input') {
+                        // 用户通过追加的 input 提交了自定义文本
+                        const customText = values?.custom_text;
+                        resolve(typeof customText === 'string' && customText.trim() ? customText.trim() : null);
+                    }
+                    else if (action === 'submit' && q.multiSelect && values) {
+                        // checker 多选提交：从 form_value 收集 checked 选项
+                        const selected = [];
+                        q.options.forEach((opt, idx) => {
+                            if (values[`opt_${idx}`] === true) {
+                                selected.push(opt.label);
                             }
-                        };
-                        sendHint().catch(() => { });
-                        permCtx.interceptNextMessage(sessionId, (msg) => {
-                            resolve(msg.content || null);
                         });
-                    }
-                    else if (q.multiSelect) {
-                        // multiSelect 按钮点击：包装为数组
-                        resolve([action]);
+                        resolve(selected.length > 0 ? selected : null);
                     }
                     else {
-                        resolve(action); // action = button key = option label
+                        resolve(action);
                     }
                 });
+                // register 已接管计数（计数 +1），现在才能安全释放 markWaiting 占位（计数 -1），避免空窗
+                if (waitMarked) {
+                    permCtx?.interactionRouter?.unmarkWaiting(sessionId);
+                    waitMarked = false;
+                }
             });
             if (answer === null) {
-                // 取消，自动选第一项
                 const firstLabel = q.options[0]?.label || '';
                 answers[q.question] = q.multiSelect ? [firstLabel] : firstLabel;
             }
@@ -320,6 +481,9 @@ export class AgentRunner {
                 answers[q.question] = answer;
             }
         }
+        if (waitMarked) {
+            permCtx?.interactionRouter?.unmarkWaiting(sessionId);
+        }
         const updatedInput = { ...input, answers };
         return { behavior: 'allow', updatedInput, decisionClassification: 'user_temporary' };
     }
@@ -395,6 +559,8 @@ export class AgentRunner {
         if (!permCtx?.channelId || !sendPrompt) {
             return { behavior: 'allow', updatedInput: input, decisionClassification: 'user_temporary' };
         }
+        // 立即暂停 idle 监控，不等卡片发完再 register
+        permCtx.interactionRouter?.markWaiting(sessionId);
         // 尝试发送交互卡片
         let cardSent = false;
         if (permCtx.adapter?.send) {
@@ -429,6 +595,7 @@ export class AgentRunner {
                         { key: 'approve', label: '✅ 批准执行', style: 'primary' },
                         { key: 'reject', label: '❌ 拒绝', style: 'danger' },
                     ],
+                    allowCustomInput: true,
                 },
                 channelId: permCtx.channelId,
                 sessionId,
@@ -455,10 +622,15 @@ export class AgentRunner {
                 logger.warn('[AgentRunner] ExitPlanMode card send failed:', err);
             }
             if (cardSent) {
+                permCtx.interactionRouter?.unmarkWaiting(sessionId);
                 return new Promise((resolve) => {
-                    permCtx.interactionRouter?.register(requestId, sessionId, (action) => {
+                    permCtx.interactionRouter?.register(requestId, sessionId, (action, values) => {
                         const trimmed = action.trim();
-                        if (trimmed === '2' || trimmed.toLowerCase() === 'reject' || trimmed === '拒绝' || trimmed === 'reject') {
+                        if (trimmed === '_custom_input') {
+                            const feedback = typeof values?.custom_text === 'string' ? values.custom_text.trim() : '';
+                            resolve({ behavior: 'deny', message: feedback || '用户提交了反馈', decisionClassification: 'user_reject' });
+                        }
+                        else if (trimmed === '2' || trimmed.toLowerCase() === 'reject' || trimmed === '拒绝') {
                             resolve({ behavior: 'deny', message: '用户拒绝了计划', decisionClassification: 'user_reject' });
                         }
                         else {
@@ -492,6 +664,7 @@ export class AgentRunner {
                 },
             };
             await sendPrompt(renderActionAsText(fallbackInteraction));
+            permCtx.interactionRouter.unmarkWaiting(sessionId);
             return new Promise((resolve) => {
                 permCtx.interactionRouter.register(fallbackRequestId, sessionId, (action) => {
                     const trimmed = action.trim();
@@ -505,6 +678,7 @@ export class AgentRunner {
             });
         }
         // 无交互能力，发提示后直接 allow
+        permCtx?.interactionRouter?.unmarkWaiting(sessionId);
         await sendPrompt('📋 计划审批\nAI 已完成规划，自动批准执行。');
         return { behavior: 'allow', updatedInput: input, decisionClassification: 'user_temporary' };
     }
@@ -512,127 +686,171 @@ export class AgentRunner {
      * SDK 原始事件 → 标准 AgentEvent 转换
      * 所有 SDK 特有的事件类型引用封装在此方法内
      */
-    async *transformStream(sdkStream, sessionId) {
+    async *transformStream(sdkStream, sessionId, callModel, callEffort, sdkModel) {
         let lastSessionId;
         // tool_use_id → tool_name 映射，用于从 SDKUserMessage 的 tool_result 块中还原工具名
         const toolUseNames = new Map();
         let turnCount = 0;
         const seenMessageIds = new Set();
-        for await (const event of sdkStream) {
-            // 提取 session_id（任意 SDK 事件都可能携带）
-            if (event.session_id && event.session_id !== lastSessionId) {
-                lastSessionId = event.session_id;
-                this.updateSessionId(sessionId, event.session_id);
-                yield { type: 'session_id', sessionId: event.session_id };
-            }
-            // system: compact_boundary → compact
-            if (event.type === 'system' && event.subtype === 'compact_boundary') {
-                yield { type: 'compact', preTokens: event.compact_metadata?.pre_tokens || 0 };
-            }
-            // system: task_progress → task_progress
-            if (event.type === 'system' && event.subtype === 'task_progress') {
-                yield {
-                    type: 'task_progress',
-                    summary: event.summary,
-                    toolUses: event.tool_uses,
-                    durationMs: event.duration_ms,
-                };
-            }
-            // system: session_state_changed → state_changed
-            if (event.type === 'system' && event.subtype === 'session_state_changed') {
-                yield { type: 'state_changed', state: event.state };
-            }
-            // assistant: 提取 tool_use 和文本（仅无 text_delta 时提取文本）
-            if (event.type === 'assistant' && event.message?.content) {
-                const msgId = event.message.id;
-                if (!msgId || !seenMessageIds.has(msgId)) {
-                    if (msgId)
-                        seenMessageIds.add(msgId);
-                    turnCount++;
+        try {
+            for await (const event of sdkStream) {
+                // 提取 session_id（任意 SDK 事件都可能携带）
+                if (event.session_id && event.session_id !== lastSessionId) {
+                    lastSessionId = event.session_id;
+                    this.updateSessionId(sessionId, event.session_id);
+                    yield { type: 'session_id', sessionId: event.session_id };
                 }
-                // 统计本轮 base agent 全部输出字符数（text + tool_use input）
-                let turnOutputChars = 0;
-                for (const content of event.message.content) {
-                    if (content.type === 'tool_use') {
-                        const inputStr = typeof content.input === 'string' ? content.input : JSON.stringify(content.input || '');
-                        turnOutputChars += inputStr.length;
-                    }
-                    else if (content.type === 'text' && content.text) {
-                        turnOutputChars += content.text.length;
-                    }
+                // system: compact_boundary → compact
+                if (event.type === 'system' && event.subtype === 'compact_boundary') {
+                    yield {
+                        type: 'compact',
+                        preTokens: event.compact_metadata?.pre_tokens || 0,
+                        postTokens: event.compact_metadata?.post_tokens,
+                        durationMs: event.compact_metadata?.duration_ms,
+                    };
                 }
-                for (const content of event.message.content) {
-                    if (content.type === 'tool_use') {
-                        if (content.id)
-                            toolUseNames.set(content.id, content.name);
-                        yield { type: 'tool_use', name: content.name, input: content.input, callId: content.id, turn: turnCount, outputTokens: turnOutputChars };
+                // system: task_progress → task_progress
+                if (event.type === 'system' && event.subtype === 'task_progress') {
+                    yield {
+                        type: 'task_progress',
+                        summary: event.summary,
+                        toolUses: event.tool_uses,
+                        durationMs: event.duration_ms,
+                    };
+                }
+                // system: session_state_changed → state_changed
+                if (event.type === 'system' && event.subtype === 'session_state_changed') {
+                    yield { type: 'state_changed', state: event.state };
+                }
+                // assistant: 提取 tool_use 和文本（仅无 text_delta 时提取文本）
+                if (event.type === 'assistant' && event.message?.content) {
+                    const msgId = event.message.id;
+                    if (!msgId || !seenMessageIds.has(msgId)) {
+                        if (msgId)
+                            seenMessageIds.add(msgId);
+                        turnCount++;
                     }
-                    else if (content.type === 'text' && content.text) {
-                        yield { type: 'text', text: content.text, outputTokens: turnOutputChars, turn: turnCount };
+                    // 统计本轮 base agent 全部输出字符数（text + tool_use input）
+                    let turnOutputChars = 0;
+                    for (const content of event.message.content) {
+                        if (content.type === 'tool_use') {
+                            const inputStr = typeof content.input === 'string' ? content.input : JSON.stringify(content.input || '');
+                            turnOutputChars += inputStr.length;
+                        }
+                        else if (content.type === 'text' && content.text) {
+                            turnOutputChars += content.text.length;
+                        }
+                    }
+                    for (const content of event.message.content) {
+                        if (content.type === 'tool_use') {
+                            if (content.id)
+                                toolUseNames.set(content.id, content.name);
+                            yield { type: 'tool_use', name: content.name, input: content.input, callId: content.id, turn: turnCount, outputTokens: turnOutputChars };
+                        }
+                        else if (content.type === 'text' && content.text) {
+                            yield { type: 'text', text: content.text, outputTokens: turnOutputChars, turn: turnCount };
+                        }
                     }
                 }
-            }
-            // user: 提取 tool_result 块（SDK 将工具结果嵌套在 SDKUserMessage 中）
-            if (event.type === 'user' && event.message?.content) {
-                const contentArray = Array.isArray(event.message.content) ? event.message.content : [];
-                for (const block of contentArray) {
-                    if (typeof block === 'object' && block !== null && block.type === 'tool_result') {
-                        const toolName = toolUseNames.get(block.tool_use_id) || '';
-                        const resultContent = typeof block.content === 'string'
-                            ? block.content
-                            : block.content != null ? JSON.stringify(block.content) : '';
-                        yield {
-                            type: 'tool_result',
-                            name: toolName,
-                            result: resultContent,
-                            isError: block.is_error === true,
-                            error: block.is_error === true ? resultContent : undefined,
-                            callId: block.tool_use_id,
-                        };
+                // user: 提取 tool_result 块（SDK 将工具结果嵌套在 SDKUserMessage 中）
+                if (event.type === 'user' && event.message?.content) {
+                    const contentArray = Array.isArray(event.message.content) ? event.message.content : [];
+                    for (const block of contentArray) {
+                        if (typeof block === 'object' && block !== null && block.type === 'tool_result') {
+                            const toolName = toolUseNames.get(block.tool_use_id) || '';
+                            const resultContent = typeof block.content === 'string'
+                                ? block.content
+                                : block.content != null ? JSON.stringify(block.content) : '';
+                            yield {
+                                type: 'tool_result',
+                                name: toolName,
+                                result: resultContent,
+                                isError: block.is_error === true,
+                                error: block.is_error === true ? resultContent : undefined,
+                                callId: block.tool_use_id,
+                            };
+                        }
                     }
                 }
-            }
-            // result → complete（含 permission_denials 提取）
-            if (event.type === 'result') {
-                // 先发出被拒绝的权限事件
-                if (Array.isArray(event.permission_denials)) {
-                    for (const denial of event.permission_denials) {
-                        yield {
-                            type: 'tool_result',
-                            name: denial.tool_name || '',
-                            result: '',
-                            isError: true,
-                            error: `权限被拒绝: ${denial.tool_name}`,
-                        };
+                // result → complete（含 permission_denials 提取）
+                if (event.type === 'result') {
+                    // 先发出被拒绝的权限事件
+                    if (Array.isArray(event.permission_denials)) {
+                        for (const denial of event.permission_denials) {
+                            yield {
+                                type: 'tool_result',
+                                name: denial.tool_name || '',
+                                result: '',
+                                isError: true,
+                                error: `权限被拒绝: ${denial.tool_name}`,
+                            };
+                        }
                     }
+                    // 剥离 SDK result 中混入的 <thinking>...</thinking> 块
+                    const cleanResult = typeof event.result === 'string'
+                        ? event.result.replace(/<thinking>[\s\S]*?<\/thinking>\s*/g, '').trim()
+                        : event.result;
+                    // 从 usage 三项求和得到当前上下文占用（与 claude-hud getTotalTokens 相同算法）
+                    const u = event.usage;
+                    const totalTokens = u
+                        ? (u.input_tokens ?? 0) + (u.cache_creation_input_tokens ?? 0) + (u.cache_read_input_tokens ?? 0)
+                        : 0;
+                    const maxTokens = sdkModel ? contextWindowFor(sdkModel) : 200000;
+                    const contextUsage = totalTokens > 0 ? {
+                        totalTokens,
+                        maxTokens,
+                        percentage: Math.round((totalTokens / maxTokens) * 100),
+                        model: callModel ?? this.model,
+                        effort: callEffort ?? this.effort,
+                    } : undefined;
+                    yield {
+                        type: 'complete',
+                        result: cleanResult,
+                        subtype: event.subtype,
+                        isError: event.is_error,
+                        errors: event.errors,
+                        durationMs: event.duration_ms,
+                        ttftMs: event.ttft_ms,
+                        costUsd: event.total_cost_usd,
+                        terminalReason: event.terminal_reason,
+                        sessionTitle: event.session_title,
+                        numTurns: event.num_turns,
+                        tokenUsage: event.usage,
+                        contextUsage,
+                    };
+                    // result 是 SDK 流的终结事件，不再等待后续（防止 interrupt 后流不关闭导致挂起）
+                    return;
                 }
-                // 剥离 SDK result 中混入的 <thinking>...</thinking> 块
-                const cleanResult = typeof event.result === 'string'
-                    ? event.result.replace(/<thinking>[\s\S]*?<\/thinking>\s*/g, '').trim()
-                    : event.result;
-                yield {
-                    type: 'complete',
-                    result: cleanResult,
-                    subtype: event.subtype,
-                    isError: event.is_error,
-                    errors: event.errors,
-                    durationMs: event.duration_ms,
-                    costUsd: event.total_cost_usd,
-                    terminalReason: event.terminal_reason,
-                    sessionTitle: event.session_title,
-                    numTurns: event.num_turns,
-                    usage: event.usage,
-                };
-                // result 是 SDK 流的终结事件，不再等待后续（防止 interrupt 后流不关闭导致挂起）
-                return;
             }
         }
+        catch (err) {
+            // 子进程崩溃（如 exited with code 1）时，把缓冲的 stderr 打出来还原真实原因。
+            // SDK 包装后的错误信息不含子进程实际报错，缓冲区才是根因所在。
+            const buf = this.recentStderr.get(sessionId);
+            if (buf && buf.length > 0) {
+                logger.error(`[AgentRunner] Subprocess stream failed (session=${sessionId}). Last ${buf.length} stderr line(s):\n${buf.join('\n')}`);
+            }
+            else {
+                logger.error(`[AgentRunner] Subprocess stream failed (session=${sessionId}) with no captured stderr.`);
+            }
+            throw err;
+        }
+        finally {
+            this.recentStderr.delete(sessionId);
+        }
     }
-    async runQuery(sessionId, prompt, projectPath, initialClaudeSessionId, images, systemPromptAppend, sessionManager) {
+    async runQuery(sessionId, prompt, projectPath, initialClaudeSessionId, images, systemPromptAppend, sessionManager, modelOverride) {
         // 记录当前 evolclaw session ID，用于 Agent ctl 环境变量注入
         this.currentEvolclawSessionId = sessionId;
         // 同步用户级配置到内存
         this.syncFromUserSettings();
+        // 异步刷新模型别名缓存（fire-and-forget，不阻塞查询）
+        if (this.baseUrl) {
+            const cached = modelAliasCache.get(this.baseUrl);
+            if (!cached || (Date.now() - cached.fetchedAt > MODEL_ALIAS_TTL_MS)) {
+                refreshModelAliases(this.baseUrl, this.apiKey);
+            }
+        }
         ensureDir(projectPath);
         ensureDir(path.join(projectPath, '.claude'));
         // 优先使用传入的 agentSessionId（从数据库恢复），否则使用内存中的
@@ -795,19 +1013,24 @@ export class AgentRunner {
         const excludeDynamic = this.config?.agents?.claude?.excludeDynamicSections === true;
         // 公共 options（新旧模式共用）
         const sdkPermissionMode = this.toSdkPermissionMode();
-        logger.info(`[AgentRunner] runQuery model=${this.model} effort=${this.effort ?? 'auto'} permMode=${this.permissionMode} sdkMode=${sdkPermissionMode}`);
+        // 本次调用使用的模型/强度：优先 modelOverride（message-processor 按 关系>agent>全局 解析后传入），
+        // 缺省回落 agent 级 this.model。作为 per-call 入参传入，无共享状态，多对端并发互不污染。
+        const callModel = modelOverride?.model || this.model;
+        const callEffort = (modelOverride?.effort ?? this.effort);
+        logger.info(`[AgentRunner] runQuery model=${callModel} effort=${callEffort ?? 'auto'} permMode=${this.permissionMode} sdkMode=${sdkPermissionMode}`);
         if (systemPromptAppend) {
             logger.info(`[AgentRunner] systemPromptAppend: ${systemPromptAppend.length} chars`);
         }
         else {
             logger.info(`[AgentRunner] systemPromptAppend: none`);
         }
+        const sdkModel = resolveSdkModel(callModel, this.baseUrl);
         const commonOptions = {
             cwd: projectPath,
-            model: this.model,
-            ...(this.effort ? { effort: this.effort } : {}),
+            model: sdkModel,
+            ...(callEffort ? { effort: callEffort } : {}),
             ...(this.claudeExecutablePath ? { pathToClaudeCodeExecutable: this.claudeExecutablePath } : {}),
-            autoCompactWindow: 200000,
+            autoCompactWindow: contextWindowFor(sdkModel),
             advisorModel: 'haiku',
             canUseTool: canUseToolCallback,
             permissionMode: sdkPermissionMode,
@@ -820,11 +1043,23 @@ export class AgentRunner {
             },
             ...(enableSummaries ? { agentProgressSummaries: true } : {}),
             stderr: (msg) => {
+                const trimmed = msg.trim();
+                if (trimmed) {
+                    // 环形缓冲：保留最近 N 行，供子进程崩溃时还原真实原因
+                    let buf = this.recentStderr.get(sessionId);
+                    if (!buf) {
+                        buf = [];
+                        this.recentStderr.set(sessionId, buf);
+                    }
+                    buf.push(trimmed);
+                    if (buf.length > AgentRunner.STDERR_BUFFER_MAX)
+                        buf.shift();
+                }
                 if (msg.includes('[ERROR]') || msg.includes('[WARN]') || msg.includes('Stream started')) {
-                    logger.info(`[Claude-stderr] ${msg.trim()}`);
+                    logger.info(`[Claude-stderr] ${trimmed}`);
                 }
                 else {
-                    logger.debug(`[Claude-stderr] ${msg.trim()}`);
+                    logger.debug(`[Claude-stderr] ${trimmed}`);
                 }
             },
             env: this.getAgentEnv()
@@ -920,7 +1155,7 @@ export class AgentRunner {
         }
         let sdkStream;
         if (images && images.length > 0) {
-            logger.debug('[AgentRunner] Creating query with images, images:', images.length);
+            logger.info('[AgentRunner] Creating query with images:', images.length, 'first image size:', images[0]?.data?.length ?? 0);
             logger.debug('[AgentRunner] Skipping resume for image message to avoid history conflict');
             const stream = new MessageStream();
             stream.push(prompt, images);
@@ -936,7 +1171,7 @@ export class AgentRunner {
             this.interruptFns.set(sessionId, () => sdkStream.interrupt());
         }
         // 返回标准 AgentEvent 流（重试由 MessageProcessor 层负责）
-        return this.transformStream(sdkStream, sessionId);
+        return this.transformStream(sdkStream, sessionId, callModel, callEffort, sdkModel);
     }
     async interrupt(sessionId) {
         const fn = this.interruptFns.get(sessionId);
@@ -961,6 +1196,7 @@ export class AgentRunner {
     cleanupStream(sessionId) {
         this.activeStreams.delete(sessionId);
         this.interruptFns.delete(sessionId);
+        this.recentStderr.delete(sessionId);
     }
     updateSessionId(sessionId, agentSessionId) {
         logger.info(`[AgentRunner] updateSessionId called: sessionId=${sessionId}, agentSessionId=${agentSessionId}`);
@@ -974,7 +1210,7 @@ export class AgentRunner {
             prompt,
             options: {
                 cwd: projectPath,
-                model: this.model,
+                model: resolveSdkModel(this.model, this.baseUrl),
                 resume: agentSessionId,
                 maxTurns: 1,
                 permissionMode: this.toSdkPermissionMode(),
@@ -1060,6 +1296,7 @@ export class AgentRunner {
                 enableFileCheckpointing: true,
                 permissionMode: this.toSdkPermissionMode(),
                 stderr: (data) => { stderrChunks.push(data); },
+                env: this.getAgentEnv(),
             }
         });
         try {