evolcore 0.0.22 → 0.0.23

package/dist/core/command/slash-handler.js

@@ -1,5 +1,4 @@
 import { hasModelSwitcher, hasPermissionController } from '../../agents/runner-types.js';
-import { getCodexEfforts } from '../../agents/codex-runner.js';
 import { normalizeBaseagent } from '../../agents/baseagent.js';
 import { buildEnvelope, isDeliveryTargetForChannel } from '../message/message-utils.js';
 import { resolvePaths, getPackageRoot, daemonControlDir } from '../../paths.js';
@@ -10,13 +9,14 @@ import path from 'path';
 import fs from 'fs';
 import { checkLatestVersion, getLocalVersion, isLinkedInstall, compareVersions } from '../../utils/npm-ops.js';
 import { loadDaemonConfig } from '../../config-store.js';
+import { remoteAccessService } from '../auth/remote-access-store.js';
 import { read as cfgRead, resolveEffective, routeFieldPath, write as cfgWrite, } from '../../config/config-manager.js';
 import { execAgentAction } from './agent-control.js';
 import { authorizeOperation, buildAuthSubject } from '../auth/auth-gateway.js';
 import { authorizationNextStep, formatAuthorizationDenial } from '../auth/authorization-denial.js';
 import { resolvePermissionMode, writeScope } from '../model/config-scope.js';
 import { formatPeerKey } from '../relation/peer-identity.js';
-import { modelMatches } from '../model/model-catalog.js';
+import { modelMatches, getCatalog, assertCatalogSelectable, resolveCatalogModel } from '../model/model-catalog.js';
 import { formatModelCheck, runModelCheck } from '../model/model-diagnostics.js';
 import { constrainResolvedModelForRole, filterModelsForRole, validateModelSelectionForRole } from '../model/model-permission.js';
 import { resolveRuntimeStringField } from '../role/runtime-policy.js';
@@ -28,7 +28,7 @@ import { isSystemControlChannel } from '../system-channels.js';
 import { spawnDetachedNode } from '../../utils/cross-platform.js';
 import { inspectDataMigrationRequirement } from '../data-migration.js';
 import { guardIdleCommand, guardKnownCommand, guardRoleCommand, guardThreadCommand, isRecognizedSlashCommand, normalizeSlashContent, probeSessionActivity, } from './slash-gate.js';
-const allEfforts = ['low', 'medium', 'high', 'xhigh', 'max'];
+const allEfforts = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'];
 const PERMISSION_MODE_KEYS = PUBLIC_PERMISSION_MODES;
 function defaultSlashChatmode(field) {
     return field === 'private' ? 'interactive' : 'proactive';
@@ -127,12 +127,8 @@ function resolveSlashRelationTarget(params) {
         role: params.role,
     };
 }
-function getAvailableEfforts(agent, model) {
-    if (agent.name === 'claude')
-        return allEfforts;
-    if (agent.name === 'codex')
-        return getCodexEfforts(model);
-    return [];
+async function getAvailableEfforts(agent, model) {
+    return (await agent.getAvailableEfforts?.(model) ?? []);
 }
 function modelDisplayLabel(agent, model) {
     const full = agent.resolveModelId?.(model);
@@ -487,7 +483,8 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
     const isProcessLevelSlash = normalizedContent === '/restart'
         || normalizedContent === '/upgrade'
         || normalizedContent === '/reload'
-        || normalizedContent.startsWith('/reload ');
+        || normalizedContent.startsWith('/reload ')
+        || normalizedContent === '/pair';
     const roleGuard = guardRoleCommand(normalizedContent, activeChatType, isAdmin || (isDaemonOwner && isProcessLevelSlash));
     if (roleGuard)
         return roleGuard;
@@ -635,7 +632,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
         const appendProcessCommands = (lines) => {
             if (!isDaemonOwner)
                 return;
-            lines.push('', '🛠️ 进程级运维：', '  /restart - 重启服务', '  /upgrade - 检查版本更新', '  /reload [aid] - 热重载 Agent 配置', '  /fa <提示词> - 单次使用宿主级 fullaccess 执行（别名: /fullaccess）');
+            lines.push('', '🛠️ 进程级运维：', '  /restart - 重启服务', '  /upgrade - 检查版本更新', '  /reload [aid] - 热重载 Agent 配置', '  /fa <提示词> - 单次使用宿主级 fullaccess 执行（别名: /fullaccess）', '  /pair - 获取远程控制配对码');
         };
         if (!isAdmin && activeChatType === 'group') {
             const lines = [
@@ -775,6 +772,23 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
         ];
         return { kind: 'command.result', text: lines.join('\n') };
     }
+    // /pair is also recognized on an Agent channel so it cannot fall through
+    // into model input. The daemon control channel has its own fast path; this
+    // branch preserves the same daemon-owner boundary for ordinary AUN chats.
+    if (normalizedContent === '/pair') {
+        if (!isDaemonOwner) {
+            return {
+                kind: 'command.error',
+                text: '❌ /pair 仅限 daemon-owner 使用；请向 daemon 控制 AID 发送 /pair。',
+            };
+        }
+        const pair = remoteAccessService.getPairingCode(userId);
+        const mins = Math.max(0, Math.round((pair.expiresAt - Date.now()) / 60000));
+        return {
+            kind: 'command.result',
+            text: `远程控制配对码：${pair.code}（约 ${mins} 分钟内有效）\n可用于 ECWeb 登录或远程 Menu 客户端授权`,
+        };
+    }
     // /evolhelp 命令：返回 JSON 格式的命令列表（供程序解析）
     if (normalizedContent === '/evolhelp') {
         const cmds = [];
@@ -835,6 +849,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
             cmds.push({ command: '/restart', description: '重启服务', category: '运维', roles: ['daemon-owner'] });
             cmds.push({ command: '/upgrade', description: '检查版本更新', category: '运维', roles: ['daemon-owner'] });
             cmds.push({ command: '/reload', args: '[aid]', description: '热重载 Agent 配置', category: '运维', roles: ['daemon-owner'] });
+            cmds.push({ command: '/pair', description: '获取远程控制配对码', category: '运维', roles: ['daemon-owner'] });
         }
         if (!isDaemonOwner && isAdmin) {
             cmds.push({ command: '/reload', description: '热重载当前 Agent 配置', category: '运维', roles: ['admin', 'owner'] });
@@ -1260,12 +1275,14 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
             catch { }
             const configuredModel = modelConfig.model;
             const configuredEffort = modelConfig.effort;
-            const rawModels = hasModelSwitcher(modelAgent) ? await modelAgent.listModels() : [];
+            const catalog = await getCatalog(modelSelector.self, fallbackBaseagent);
+            assertCatalogSelectable(catalog);
+            const rawModels = catalog.models.map(m => m.id);
             const models = hasModelSwitcher(modelAgent)
-                ? filterModelsForRole(identity.role, fallbackBaseagent, rawModels, modelAgent.resolveModelId?.bind(modelAgent))
+                ? filterModelsForRole(identity.role, fallbackBaseagent, rawModels, model => resolveCatalogModel(catalog, model), modelSelector.self)
                 : rawModels;
             const currentModel = hasModelSwitcher(modelAgent) ? (configuredModel || modelAgent.getModel()) : modelAgent.name;
-            const efforts = getAvailableEfforts(modelAgent, currentModel);
+            const efforts = await getAvailableEfforts(modelAgent, currentModel);
             const currentEffort = configuredEffort ?? modelAgent.getEffort?.() ?? 'auto';
             const canReadModelList = await canReadSlashModelList({
                 subject: authSubject,
@@ -1290,7 +1307,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
                         kind: 'command-card',
                         title: '🤖 切换模型',
                         buttons: models.map((m) => {
-                            const display = modelDisplayLabel(modelAgent, m);
+                            const display = modelDisplayLabel(modelAgent, m) + (catalog.models.find(e => e.id === m)?.verification === 'unverified' ? ' [unverified]' : '');
                             return {
                                 label: modelMatches(modelAgent, m, currentModel) ? `✓ ${display}` : display,
                                 command: `/model ${m}`,
@@ -1307,6 +1324,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
                 return { kind: 'command.result', text: cardResult };
             }
             // 降级：文本
+            const catalogHint = `[${catalog.source}/${catalog.freshness}] ${catalog.diagnostics.map(d => d.message).join('; ')}`;
             const modelList = models.map((m) => `  ${modelMatches(modelAgent, m, currentModel) ? '✓' : ' '} ${modelDisplayLabel(modelAgent, m)}`).join('\n');
             const effortHint = efforts.length > 0
                 ? `\n推理强度: ${currentEffort === 'auto' ? 'auto (SDK默认)' : currentEffort}  (使用 /effort 调整)`
@@ -1322,8 +1340,8 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
                 const usage = !args && isAdmin ? '\n\n用法: /model <模型>' : '';
                 return {
                     kind: 'command.result',
-                    text: `当前模型: ${modelDisplayLabel(modelAgent, currentModel)}${effortHint}\n\n可用模型：\n${modelList}${usage}`,
-                    ...(args === 'list' ? { structured: { ok: true, model: currentModel, effort: currentEffort, baseagent: fallbackBaseagent, models } } : {}),
+                    text: `当前模型: ${modelDisplayLabel(modelAgent, currentModel)}${effortHint}${rawModels.includes(currentModel) ? '' : ' [当前配置，未列入目录]'}\n${catalogHint}\n\n可用模型：\n${modelList}${usage}`,
+                    ...(args === 'list' ? { structured: { ok: true, model: currentModel, effort: currentEffort, baseagent: fallbackBaseagent, models, revision: catalog.revision, source: catalog.source, freshness: catalog.freshness, diagnostics: catalog.diagnostics } } : {}),
                 };
             }
             return { kind: 'command.result', text: `当前模型: ${modelDisplayLabel(modelAgent, currentModel)}${effortHint}` };
@@ -1384,7 +1402,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
         catch { }
         const models = hasModelSwitcher(modelAgent) ? await modelAgent.listModels() : [];
         const allowedModels = hasModelSwitcher(modelAgent)
-            ? filterModelsForRole(identity.role, fallbackBaseagent, models, modelAgent.resolveModelId?.bind(modelAgent))
+            ? filterModelsForRole(identity.role, fallbackBaseagent, models, modelAgent.resolveModelId?.bind(modelAgent), selfAID ?? this.resolveSelfAID(channel))
             : models;
         const parts = args.split(/\s+/);
         let newModel;
@@ -1392,7 +1410,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
         if (parts.length === 1) {
             const arg = parts[0];
             const currentModel = hasModelSwitcher(modelAgent) ? (configuredModel || modelAgent.getModel()) : modelAgent.name;
-            const efforts = getAvailableEfforts(modelAgent, currentModel);
+            const efforts = await getAvailableEfforts(modelAgent, currentModel);
             // effort 相关参数统一转发到 /effort
             if (efforts.includes(arg) || arg === 'auto') {
                 const delegated = await this.handle(`/effort ${arg}`, channel, channelId, undefined, userId, threadId);
@@ -1425,7 +1443,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
             if (!models.includes(modelArg)) {
                 return { kind: 'command.error', text: `❌ 无效的模型ID: ${modelArgRaw}` };
             }
-            const targetEfforts = getAvailableEfforts(modelAgent, modelArg);
+            const targetEfforts = await getAvailableEfforts(modelAgent, modelArg);
             if (targetEfforts.length === 0) {
                 return { kind: 'command.error', text: `⚠️ ${modelArg} 不支持推理强度设置` };
             }
@@ -1583,7 +1601,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
             const configuredModel = modelConfig.model;
             const configuredEffort = modelConfig.effort;
             const currentModel = hasModelSwitcher(effortAgent) ? (configuredModel || effortAgent.getModel()) : effortAgent.name;
-            const efforts = getAvailableEfforts(effortAgent, currentModel);
+            const efforts = await getAvailableEfforts(effortAgent, currentModel);
             const currentEffort = configuredEffort ?? effortAgent.getEffort?.() ?? 'auto';
             if (efforts.length === 0) {
                 return { kind: 'command.error', text: '⚠️ 当前模型不支持推理强度设置' };
@@ -1645,7 +1663,7 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
         const configuredModel = modelConfig.model;
         const configuredEffort = modelConfig.effort;
         const currentModel = hasModelSwitcher(effortAgent) ? (configuredModel || effortAgent.getModel()) : effortAgent.name;
-        const efforts = getAvailableEfforts(effortAgent, currentModel);
+        const efforts = await getAvailableEfforts(effortAgent, currentModel);
         const currentEffort = configuredEffort ?? effortAgent.getEffort?.() ?? 'auto';
         if (efforts.length === 0) {
             return { kind: 'command.error', text: '⚠️ 当前模型不支持推理强度设置' };
@@ -2502,6 +2520,8 @@ export async function handleSlashCommand(content, channel, channelId, sendMessag
         if (session && previousAgent) {
             // Reset agent backend state so the new
             // session starts with a fresh conversation history
+            // Clearing is a pure backend lifecycle operation; model selection is
+            // resolved only when the new session performs its next turn.
             await previousAgent.clearSession(session.id, session.agentSessionId || '', session.projectPath);
             await previousAgent.closeSession(session.id);
         }

package/dist/core/message/message-queue.js

@@ -234,6 +234,7 @@ export class MessageQueue {
         }
         const tmp = `${this.dedupPath}.tmp-${process.pid}-${Date.now()}-${this.enqueueSequence}`;
         let fd;
+        let renamed = false;
         try {
             fs.mkdirSync(path.dirname(this.dedupPath), { recursive: true });
             const content = Buffer.from(JSON.stringify({
@@ -253,20 +254,22 @@ export class MessageQueue {
             fs.closeSync(fd);
             fd = undefined;
             fs.renameSync(tmp, this.dedupPath);
-            const persistedFd = fs.openSync(this.dedupPath, 'r');
-            try {
-                fs.fsyncSync(persistedFd);
-            }
-            finally {
-                fs.closeSync(persistedFd);
-            }
+            renamed = true;
             if (process.platform !== 'win32') {
-                const directoryFd = fs.openSync(path.dirname(this.dedupPath), 'r');
                 try {
-                    fs.fsyncSync(directoryFd);
+                    const directoryFd = fs.openSync(path.dirname(this.dedupPath), 'r');
+                    try {
+                        fs.fsyncSync(directoryFd);
+                    }
+                    finally {
+                        fs.closeSync(directoryFd);
+                    }
                 }
-                finally {
-                    fs.closeSync(directoryFd);
+                catch (error) {
+                    // The file contents were already fsynced before rename. A platform
+                    // that cannot fsync directory metadata must not make a committed
+                    // dedup claim look uncommitted to the caller.
+                    logger.warn(`[Queue] Durable dedup ledger directory fsync unavailable after commit: ${error instanceof Error ? error.message : String(error)}`);
                 }
             }
             return true;
@@ -279,11 +282,13 @@ export class MessageQueue {
                 }
                 catch { }
             }
-            try {
-                fs.unlinkSync(tmp);
+            if (!renamed) {
+                try {
+                    fs.unlinkSync(tmp);
+                }
+                catch { }
             }
-            catch { }
-            return false;
+            return renamed;
         }
     }
     hasEnqueuedMessage(sessionKey, messageId, selfAID) {

package/dist/core/message/response-engine.js

@@ -325,44 +325,71 @@ function withFallbackModel(override, candidate) {
         ...(candidate.effort ? { effort: candidate.effort } : { effortMode: 'model_default' }),
     };
 }
+/**
+ * Keep the per-task fullaccess fact in the user-message prompt rather than in
+ * the long-lived system/developer prompt. This avoids changing a runner's
+ * system-prompt cache while still giving the model an explicit, auditable
+ * execution-mode signal for the privileged task.
+ */
+const FULL_ACCESS_MESSAGE_PROMPT_CONTEXT = '【当前权限模式提升为fullaccess，可不受限制执行命令，仅当前轮 task 有效】';
+/**
+ * Put the one-task permission notice between the rendered message envelope and
+ * the user's literal input. The normal message fragment ends its envelope in
+ * `›`; matching the dated normal-message form avoids placing the notice inside
+ * an owner-hint envelope when one is present before the real inbound message.
+ */
+function insertMessagePromptContext(prompt, context, userContent) {
+    if (!context)
+        return prompt;
+    const headers = [...prompt.matchAll(/‹\d{4}-\d{2}-\d{2} [^\n]*· from:[^\n]*›/gu)];
+    const renderedHeader = headers[0];
+    const headerEnd = renderedHeader
+        ? (renderedHeader.index ?? 0) + renderedHeader[0].length
+        : -1;
+    if (headerEnd >= 0) {
+        const remainder = prompt.slice(headerEnd).replace(/^\n+/u, '');
+        return `${prompt.slice(0, headerEnd)}\n${context}${remainder ? `\n${remainder}` : ''}`;
+    }
+    if (userContent?.trim()) {
+        const contentStart = prompt.indexOf(userContent);
+        if (contentStart >= 0) {
+            return `${prompt.slice(0, contentStart)}${context}\n${prompt.slice(contentStart)}`;
+        }
+    }
+    // Internal recovery prompts have no message envelope; keep their context at
+    // the front so the runner still sees it for the same task.
+    return `${context}\n\n${prompt}`;
+}
 function getStreamErrorText(result) {
-    return [
+    const parts = [
         result.errors?.join('\n'),
-        result.lastReplyText,
-        result.fullText,
         result.subtype,
         result.terminalReason,
     ]
-        .filter((part) => typeof part === 'string' && part.trim().length > 0)
-        .join('\n');
+        .filter((part) => typeof part === 'string' && part.trim().length > 0);
+    // Some older runners put the terminal diagnostic in `result`, but never
+    // treat streamed assistant text as an error merely because the turn failed.
+    const fullText = result.fullText.trim();
+    const fullTextLooksLikeDiagnostic = fullText.length > 0 && (isRetryableError(new Error(fullText))
+        || isContextTooLongText(fullText));
+    if (result.isError && !result.hasReceivedText && fullText
+        && (parts.length === 0 || fullTextLooksLikeDiagnostic)) {
+        parts.unshift(fullText);
+    }
+    return parts.join('\n');
 }
 function getStreamErrorMessage(result, includeEmoji = false) {
     const raw = getStreamErrorText(result) || '任务执行失败';
     const mapped = getErrorMessage(new Error(raw), result.terminalReason, includeEmoji);
     const generic = includeEmoji ? '❌ 处理消息时出错，请稍后重试' : '处理消息时出错，请稍后重试';
-    if (!result.terminalReason && raw && mapped === generic) {
+    // Unknown runner-specific terminal reasons do not have a friendly mapping.
+    // Preserve their diagnostic text instead of collapsing it to the generic
+    // "处理消息时出错" message.
+    if (raw && mapped === generic) {
         return includeEmoji ? `❌ ${raw}` : raw;
     }
     return mapped;
 }
-function getRuntimeErrorMessage(raw, includeEmoji = false) {
-    const fallback = raw || '任务执行失败';
-    const mapped = getErrorMessage(new Error(fallback), undefined, includeEmoji);
-    const generic = includeEmoji ? '❌ 处理消息时出错，请稍后重试' : '处理消息时出错，请稍后重试';
-    if (mapped === generic)
-        return includeEmoji ? `❌ ${fallback}` : fallback;
-    return mapped;
-}
-function normalizeComparableText(text) {
-    return text.replace(/\s+/g, ' ').trim();
-}
-function isPendingTextSameAsStreamError(pendingText, errorText) {
-    const pending = normalizeComparableText(pendingText);
-    const error = normalizeComparableText(errorText);
-    if (!pending || !error || pending.length < 20)
-        return false;
-    return pending === error || pending.includes(error) || error.includes(pending);
-}
 function streamHitContextLimit(result) {
     return result.terminalReason === 'prompt_too_long' ||
         isContextTooLongText(result.lastReplyText) ||
@@ -446,12 +473,6 @@ function getContextTooLongHint(agent) {
     }
     return '上下文过长，请精简提问，或使用 /new 新建会话后继续';
 }
-function getContextCompactFailedHint(agent) {
-    if (canCompactAgent(agent)) {
-        return '上下文过长，自动压缩失败，请手动输入 /compact 重试';
-    }
-    return '上下文过长，请精简提问，或使用 /new 新建会话后继续';
-}
 function canCompactAgent(agent) {
     return hasCompact(agent) && agent.capabilities?.compact !== false;
 }
@@ -1042,7 +1063,7 @@ export class ResponseEngine {
         return true;
     }
     async retryAfterContextRecovery(prompt, opts) {
-        const { streamKey, renderer, agent, session, absoluteProjectPath, effectiveSystemPrompt, modelOverride, runtimeEnv, resetTimer, shouldSuppress, proactive, turnLease, beforeRunQuery, } = opts;
+        const { streamKey, renderer, agent, session, absoluteProjectPath, effectiveSystemPrompt, modelOverride, runtimeEnv, resetTimer, shouldSuppress, proactive, turnLease, beforeRunQuery, messagePromptContext, } = opts;
         if (!session.agentSessionId || !canCompactAgent(agent)) {
             throw new Error('CONTEXT_COMPACT_FAILED');
         }
@@ -1081,7 +1102,7 @@ export class ResponseEngine {
             await this.emitOperationalNotice(renderer, '✅ 压缩完成，继续处理...', 'info', 'compact-retry');
             await renderer.flush();
             beforeRunQuery?.();
-            const retryStream = await agent.runQuery(session.id, prompt, absoluteProjectPath, session.agentSessionId, undefined, effectiveSystemPrompt, this.sessionManager, recoveryModelOverride, runtimeEnv);
+            const retryStream = await agent.runQuery(session.id, insertMessagePromptContext(prompt, messagePromptContext), absoluteProjectPath, session.agentSessionId, undefined, effectiveSystemPrompt, this.sessionManager, recoveryModelOverride, runtimeEnv);
             if (await this.cancelStaleRunnerStart(agent, session.id, streamKey, turnLease, turnLease.taskId)) {
                 throw new Error('RUNNER_START_SUPERSEDED');
             }
@@ -1365,11 +1386,11 @@ export class ResponseEngine {
         '/model', '/effort', '/agent', '/slist', '/session', '/rename', '/repair', '/fork',
         '/stop', '/pause', '/resume', '/clear', '/compact', '/renew', '/del', '/perm', '/file', '/check',
         '/s ', '/name ', '/rewind', '/rw', '/rw ', '/activity', '/observable', '/chatmode',
-        '/aid', '/upgrade', '/evolagent',
+        '/aid', '/upgrade', '/evolagent', '/pair ',
     ];
     /** 判断消息内容是否为已知命令 */
     isKnownCommand(content) {
-        return content === '/s' ||
+        return content === '/s' || content === '/pair' ||
             ResponseEngine.COMMAND_PREFIXES.some(cmd => content.startsWith(cmd));
     }
     /**
@@ -2689,6 +2710,7 @@ export class ResponseEngine {
                 : body;
             // 先用裸文本兜底；vars 构造完成后用消息渲染层重算（见下方 effectivePrompt 重赋值）。
             let effectivePrompt = wrapPrompt(message.content);
+            let fullAccessMessagePromptContext;
             let renderResult;
             let effectiveSystemPrompt;
             let modelOverride;
@@ -2784,6 +2806,9 @@ export class ResponseEngine {
                 // records all observe the same normalized value.
                 permissionContext.permissionMode = effectivePermissionMode;
                 permissionContext.executionPermission = fullAccessAuthorization;
+                fullAccessMessagePromptContext = fullAccessAuthorization
+                    ? FULL_ACCESS_MESSAGE_PROMPT_CONTEXT
+                    : undefined;
                 // 按 关系级 > agent级 > 全局 解析本次调用的模型/强度，作为 per-call 入参传入 runQuery。
                 // 不缓存、不绑会话——改关系级/agent级后该范围所有会话的下条消息即时生效；
                 // 多对端并发各自独立解析、各自传参，无共享状态可被污染。
@@ -2878,7 +2903,9 @@ export class ResponseEngine {
                 agentModel = (typeof agent.getModel === 'function') ? agent.getModel() : undefined;
                 primaryModel = modelOverride?.model || agentModel;
                 if (!hasExplicitTriggerModel && primaryModel) {
-                    const fallbackResolution = eligibleFallbackModels({
+                    const fallbackResolution = await eligibleFallbackModels({
+                        selfAid: selfAid || undefined,
+                        catalog: agent.getModelCatalog?.bind(agent),
                         baseagent: normalizedBaseagent.canonical,
                         currentModel: primaryModel,
                         resolveModelId: typeof agent.resolveModelId === 'function'
@@ -3218,6 +3245,12 @@ export class ResponseEngine {
                     if (v2HandoffIds.length > 0 && renderResult?.body.trim())
                         handoffPromptRendered = true;
                 }
+                // Keep this notice in the per-turn message prompt. Injecting it into
+                // the system prompt would alter long-lived runner prompt caches and
+                // could make a one-turn grant look persistent.
+                if (fullAccessMessagePromptContext && effectivePrompt.trim()) {
+                    effectivePrompt = insertMessagePromptContext(effectivePrompt, fullAccessMessagePromptContext, message.content);
+                }
                 // 空消息防护：在 agent 调用之前检查 prompt 是否为空
                 // 防止空消息（或纯空格消息）浪费 API 调用
                 if (!effectivePrompt.trim()) {
@@ -3527,6 +3560,7 @@ export class ResponseEngine {
                         proactiveSelfAid: selfAid,
                         permissionMode: effectivePermissionMode,
                         turnLease,
+                        messagePromptContext: fullAccessMessagePromptContext,
                         beforeRunQuery: fullAccessAuthorization
                             ? () => this.assertCurrentFullAccessExecutionAuthorization(fullAccessAuthorization, agent)
                             : undefined,
@@ -3573,6 +3607,7 @@ export class ResponseEngine {
                     proactiveSelfAid: selfAid,
                     permissionMode: effectivePermissionMode,
                     turnLease,
+                    messagePromptContext: fullAccessMessagePromptContext,
                     beforeRunQuery: fullAccessAuthorization
                         ? () => this.assertCurrentFullAccessExecutionAuthorization(fullAccessAuthorization, agent)
                         : undefined,
@@ -5254,8 +5289,10 @@ export class ResponseEngine {
                     // 运行时错误（Codex: turn.failed / item error）
                     if (event.type === 'error') {
                         logger.warn(`[ResponseEngine] error event: ${event.errorType}: ${event.error}`);
-                        // 记录错误文本到 lastReplyText，供后续 isPromptTooLong 检测
-                        lastReplyText += event.error || '';
+                        // Keep runtime diagnostics separate from assistant prose.  Appending
+                        // an error to lastReplyText makes the terminal flush treat the error
+                        // as user-visible model output and can also make it look like a
+                        // duplicate of a pending text event.
                         completeResult = {
                             ...completeResult,
                             isError: true,
@@ -5265,14 +5302,6 @@ export class ResponseEngine {
                             lastReplyText,
                             fullText: '',
                         };
-                        // 上下文过长的错误不在此处输出 notice，留给外层 isPromptTooLong 触发 auto-compact
-                        const isContextError = isContextTooLongText(event.error || '');
-                        const isRetryableRuntimeError = isRetryableError(new Error(event.error || ''));
-                        const isRecoverableModelError = classifyError(new Error(event.error || '')) === ErrorType.MODEL_UNAVAILABLE;
-                        if (!isContextError && !isRetryableRuntimeError && !isRecoverableModelError && !hasErrorResult && !shouldSuppress()) {
-                            hasErrorResult = true;
-                            renderer.addNotice(getRuntimeErrorMessage(event.error || '任务执行失败', false), 'warn', 'runtime-error', true);
-                        }
                     }
                     // 完成事件
                     // SDK 可能产生多个 complete 事件（如 subagent 或 auto-compact 二次查询），
@@ -5291,34 +5320,6 @@ export class ResponseEngine {
                             madeProgress = true;
                         // thought jsonl 写入已下沉到 aun.ts:sendThought 成功后，
                         // 由那里按 LLM 输出的每个 text item 单独写一条，此处不再写。
-                        // 失败且无前置错误输出：显示 errors 摘要
-                        // 但用户主动中断（新消息打断 或 /stop 命令）时不显示错误提示
-                        // 上下文过长的错误留给外层 isPromptTooLong 触发 auto-compact，不在此处输出
-                        const interruptReason = this.interruptedSessions.get(session.id);
-                        const isUserInterrupt = isNonErrorInterruptReason(interruptReason)
-                            || isExpectedTerminalInterrupt(event.terminalReason);
-                        const isContextTooLong = event.terminalReason === 'prompt_too_long'
-                            || isContextTooLongText(event.errors?.join(' ') || '')
-                            || isContextTooLongText(lastReplyText);
-                        const completeErrorText = getStreamErrorText(completeResult);
-                        const isRetryableCompleteError = event.isError && !isContextTooLong && completeErrorText
-                            ? isRetryableError(new Error(completeErrorText))
-                            : false;
-                        const isRecoverableModelError = event.isError && !isContextTooLong && completeErrorText
-                            ? classifyError(new Error(completeErrorText)) === ErrorType.MODEL_UNAVAILABLE
-                            : false;
-                        if (event.isError && completeErrorText && isPendingTextSameAsStreamError(renderer.getRemainingText(), completeErrorText)) {
-                            renderer.discardPendingText();
-                        }
-                        if (event.isError && !hasErrorResult && !shouldSuppress() && !isUserInterrupt && !isContextTooLong && !isRetryableCompleteError && !isRecoverableModelError) {
-                            // 使用 terminalReason 提供更友好的错误提示（不带 emoji，由 formatter 统一加）
-                            const userFriendlyMessage = event.terminalReason === 'prompt_too_long'
-                                ? getContextTooLongHint(agent)
-                                : event.terminalReason === 'context_compact_failed'
-                                    ? getContextCompactFailedHint(agent)
-                                    : getStreamErrorMessage(completeResult, false);
-                            renderer.addNotice(userFriendlyMessage, 'warn', 'task-error', true);
-                        }
                         // 中间 complete：flush 掉已有 activities（不带 isFinal），让中间结果及时显示
                         // 最终文本留给流结束后的统一 flush(true)
                         if (renderer.hasContent()) {