evolcore 0.0.21 → 0.0.23

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

@@ -37,7 +37,7 @@ import { resolveAgentLifecycle } from '../../config/lifecycle.js';
 import { authorizationConfigRevision, checkRoleAccess, getFirstStaticAgentOwner, listStaticAgentAdmins, listStaticAgentOwners, resolvePeerRoleDetail, roleToSessionIdentity, } from '../../config/peer-role-resolver.js';
 import { insertUsageEvent, insertContextBreakdown, insertModelCalls } from '../../stats/writer.js';
 import { normalizeUsage } from '../../stats/normalizer.js';
-import { resolvePrices } from '../../stats/price-resolver.js';
+import { resolvePrices, roundCostForOutput } from '../../stats/price-resolver.js';
 import { getBudgetStatus } from '../../stats/budget.js';
 import { formatUsageSubjectKey, getRoleBudgetStatus } from '../../stats/role-budget.js';
 import { snapshot } from './response-snapshot.js';
@@ -47,7 +47,7 @@ import { registerBuiltinModes } from '../../response-system/modes/index.js';
 import { deriveSessionTitle, shouldAutoFillSessionTitle } from '../session/session-title.js';
 import { SessionTurnCoordinator } from '../session/session-turn-coordinator.js';
 import { recordTriggerExecutionAnomaly } from '../../trigger/anomaly-store.js';
-import { classifyToolErrorCode } from '../permission/tool-error-code.js';
+import { classifyToolErrorCode, normalizeToolErrorCode } from '../permission/tool-error-code.js';
 import { normalizeExecutionPermissionMode } from '../permission/mode.js';
 import { buildToolLifecycleEventKey } from '../audit/event-key.js';
 import { isFullAccessEnabled, loadDaemonConfig } from '../../config-store.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;
 }
@@ -637,20 +658,44 @@ export class ResponseEngine {
         // 监听中断事件，标记被中断的 session
         this.eventBus.subscribe('task:interrupted', (event) => {
             if ('sessionId' in event && event.sessionId) {
+                const eventIdentity = event;
+                const eventTaskId = eventIdentity.taskId;
+                const activeTask = this.activeTaskSpans.get(event.sessionId);
+                // A delayed interruption from an older task must never invalidate a
+                // newer task that has already claimed the same session.
+                if (eventTaskId && activeTask && activeTask.taskId !== eventTaskId)
+                    return;
                 const reason = (event.reason || 'new_message');
-                this.turnCoordinator.invalidate(event.sessionId, reason);
+                const currentTurn = this.turnCoordinator.invalidate(event.sessionId, reason, { taskId: eventTaskId, generation: eventIdentity.generation });
+                if ((eventTaskId || eventIdentity.generation !== undefined) && !currentTurn)
+                    return;
                 this.interruptedSessions.set(event.sessionId, reason);
                 this.cancelRetryDelays(event.sessionId);
-                this.agentDelegationRegistry?.revokeSession(event.sessionId);
+                // The event can be delayed relative to a newer task on the same
+                // session. Revoke only when the producer supplied the matching task;
+                // the interrupt path below performs the authoritative task-scoped
+                // revocation for legacy events without a taskId.
+                const interruptedTaskId = eventTaskId;
+                if (interruptedTaskId && activeTask?.taskId === interruptedTaskId) {
+                    this.agentDelegationRegistry?.revokeTask(event.sessionId, interruptedTaskId);
+                }
             }
         });
         this.eventBus.subscribe('task:completed', event => {
-            if ('sessionId' in event && event.sessionId)
-                this.activeTaskSpans.delete(event.sessionId);
+            if ('sessionId' in event && event.sessionId) {
+                const taskId = event.taskId;
+                const active = this.activeTaskSpans.get(event.sessionId);
+                if (!taskId || active?.taskId === taskId)
+                    this.activeTaskSpans.delete(event.sessionId);
+            }
         });
         this.eventBus.subscribe('task:error', event => {
-            if ('sessionId' in event && event.sessionId)
-                this.activeTaskSpans.delete(event.sessionId);
+            if ('sessionId' in event && event.sessionId) {
+                const taskId = event.taskId;
+                const active = this.activeTaskSpans.get(event.sessionId);
+                if (!taskId || active?.taskId === taskId)
+                    this.activeTaskSpans.delete(event.sessionId);
+            }
         });
         // 初始化响应模式协调器，注册内置模式（interactive/proactive）
         const registry = new ResponseModeRegistry();
@@ -1018,7 +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');
         }
@@ -1057,7 +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');
             }
@@ -1274,7 +1319,7 @@ export class ResponseEngine {
             causation: opts.causation ?? message.causation,
         });
     }
-    publishTriggerExecutionSkipped(message, reason, causation) {
+    publishTriggerExecutionSkipped(message, reason, causation, interruption) {
         const terminal = this.claimTriggerTerminal(message);
         if (!terminal)
             return;
@@ -1287,6 +1332,7 @@ export class ResponseEngine {
             attemptId: trigger.attemptId,
             originTriggerId: trigger.triggerId,
             reason,
+            ...(interruption ?? {}),
             targetChannel: message.channel,
             targetChannelId: message.channelId,
             fireTime: trigger.fireTime,
@@ -1294,16 +1340,25 @@ export class ResponseEngine {
         });
     }
     /** Close the internal daemon conversation before returning from an interrupted Trigger turn. */
-    async publishTriggerExecutionInterrupted(message, adapter, envelope, reason, causation) {
+    async publishTriggerExecutionInterrupted(message, adapter, envelope, reason, causation, generation) {
+        const interruption = reason === 'stale_generation'
+            ? {
+                reasonCode: 'stale_generation',
+                decisionSource: 'infrastructure',
+                executionState: 'interrupted',
+                ...(generation?.expected !== undefined ? { generation: generation.expected } : {}),
+                ...(generation?.current !== undefined ? { currentGeneration: generation.current } : {}),
+            }
+            : undefined;
         if (this.isTrustedDaemonTrigger(message)) {
             await adapter.send(envelope, {
                 kind: 'status.interrupted',
-                metadata: { reason },
+                metadata: { reason, ...(interruption ?? {}) },
             }).catch(error => {
                 logger.warn(`[ResponseEngine] Failed to close interrupted Trigger run=${message.triggerMeta?.runId ?? '<unknown>'}: ${error instanceof Error ? error.message : String(error)}`);
             });
         }
-        this.publishTriggerExecutionSkipped(message, reason, causation);
+        this.publishTriggerExecutionSkipped(message, reason, causation, interruption);
     }
     publishTriggerExecutionCompleted(message, messageId, durationMs, causation) {
         const terminal = this.claimTriggerTerminal(message);
@@ -1331,11 +1386,11 @@ export class ResponseEngine {
         '/model', '/effort', '/agent', '/slist', '/session', '/rename', '/repair', '/fork',
         '/stop', '/pause', '/resume', '/clear', '/compact', '/renew', '/del', '/perm', '/file', '/check',
         '/s ', '/name ', '/rewind', '/rw', '/rw ', '/activity', '/observable', '/chatmode',
-        '/aid', '/upgrade', '/evolagent',
+        '/aid', '/upgrade', '/evolagent', '/pair ',
     ];
     /** 判断消息内容是否为已知命令 */
     isKnownCommand(content) {
-        return content === '/s' ||
+        return content === '/s' || content === '/pair' ||
             ResponseEngine.COMMAND_PREFIXES.some(cmd => content.startsWith(cmd));
     }
     /**
@@ -2130,6 +2185,8 @@ export class ResponseEngine {
                 this.eventBus.publish({
                     type: 'task:error',
                     sessionId: session.id,
+                    taskId,
+                    generation: turnLease?.generation,
                     error: timeoutError.message,
                     errorType,
                     agentName: agentNameForStats,
@@ -2255,7 +2312,10 @@ export class ResponseEngine {
             const contentPreview = formatInboundMessageLogText(message.content);
             logger.info(`[${message.channel}] ${message.channelId}: ${contentPreview}${imageInfo}${modeInfo}${e2eeInfo}`);
             // 构建 peer 标识（优先 peerName，退化到 peerId / channelId）
-            const peerName = session.metadata?.peerName ?? message.peerName;
+            // Group sessions are shared by multiple senders. The current inbound
+            // message is authoritative; session metadata is only a fallback for
+            // legacy paths that do not carry a per-message display name.
+            const peerName = message.peerName ?? session.metadata?.peerName;
             const peerId = session.metadata?.peerId ?? message.peerId ?? message.channelId;
             const peerShort = peerId ? peerId.split('.')[0].split(':')[0] : '?';
             const peerLabel = peerName && peerName !== peerShort ? `${peerShort}(${peerName})` : peerShort;
@@ -2266,7 +2326,16 @@ export class ResponseEngine {
             turnLease = await this.turnCoordinator.begin(session, taskId);
             // 记录开始处理
             const taskEncrypt = message.replyContext?.metadata?.encrypted != null ? !!(message.replyContext.metadata.encrypted) : undefined;
-            this.eventBus.publish({ type: 'task:started', sessionId: session.id, agentName: agentNameForStats, encrypt: taskEncrypt, chatmode, causation: taskCausation });
+            this.eventBus.publish({
+                type: 'task:started',
+                sessionId: session.id,
+                taskId,
+                generation: turnLease.generation,
+                agentName: agentNameForStats,
+                encrypt: taskEncrypt,
+                chatmode,
+                causation: taskCausation,
+            });
             this.touchAgentActivity(channelKey);
             // Upgrade the channel acknowledgement at task start, before compaction and
             // runner setup, so processing feedback remains visible for the whole run.
@@ -2292,6 +2361,7 @@ export class ResponseEngine {
                 adapter,
                 envelope,
                 agentAid: session.selfAID,
+                permissionMode: () => effectivePermissionMode,
                 flushDelay: (options?.flushDelay ?? this.agentRegistry?.resolveByChannel(channelKey)?.config?.flush_delay ?? 3) * 1000,
                 suppressActivityItems: isProactive ? false : middleOutputMode !== 'all',
                 suppressIntermediateText: isProactive ? false : middleOutputMode === 'none',
@@ -2640,6 +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;
@@ -2735,6 +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。
@@ -2829,7 +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'
@@ -3108,7 +3184,7 @@ export class ResponseEngine {
                             }];
                     const peerItems = (() => {
                         if (message.handoffDelivery && this.handoffRuntime) {
-                            const items = this.handoffRuntime.buildPromptItems(message);
+                            const items = this.handoffRuntime.buildPromptItems(message, peerRole);
                             if (items.length > 0) {
                                 v2HandoffIds = Array.from(new Set(items.flatMap(item => (item.handoff?.handoffIds?.length
                                     ? item.handoff.handoffIds
@@ -3169,6 +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()) {
@@ -3205,12 +3287,14 @@ export class ResponseEngine {
                     peerName: peerName || undefined,
                     peerType: message.peerType || session.metadata?.peerType || undefined,
                     peerRole,
+                    permissionMode: effectivePermissionMode,
                     ...(fullAccessAuthorization ? {
                         processRole: fullAccessAuthorization.processRole,
                         dataScope: fullAccessAuthorization.dataScope,
                         authorizedBy: fullAccessAuthorization.authorizedBy,
                         executionSource: fullAccessAuthorization.source,
                     } : {}),
+                    daemonRuntimeEpoch: this.agentDelegationRegistry?.getRuntimeEpoch(),
                     threadId: session.threadId || undefined,
                     sessionRuntimeDir,
                     runtimeLockDir: ensureRuntimeLockDir(sessionRuntimeDir),
@@ -3476,6 +3560,7 @@ export class ResponseEngine {
                         proactiveSelfAid: selfAid,
                         permissionMode: effectivePermissionMode,
                         turnLease,
+                        messagePromptContext: fullAccessMessagePromptContext,
                         beforeRunQuery: fullAccessAuthorization
                             ? () => this.assertCurrentFullAccessExecutionAuthorization(fullAccessAuthorization, agent)
                             : undefined,
@@ -3522,6 +3607,7 @@ export class ResponseEngine {
                     proactiveSelfAid: selfAid,
                     permissionMode: effectivePermissionMode,
                     turnLease,
+                    messagePromptContext: fullAccessMessagePromptContext,
                     beforeRunQuery: fullAccessAuthorization
                         ? () => this.assertCurrentFullAccessExecutionAuthorization(fullAccessAuthorization, agent)
                         : undefined,
@@ -3547,7 +3633,10 @@ export class ResponseEngine {
                     }
                     this.agentDelegationRegistry?.revokeTask(session.id, taskId);
                     logger.info(`[ResponseEngine] Stale turn stopped before publish: session=${session.id} task=${taskId}`);
-                    await this.publishTriggerExecutionInterrupted(message, adapter, envelope, 'stale_generation', taskCausation);
+                    await this.publishTriggerExecutionInterrupted(message, adapter, envelope, 'stale_generation', taskCausation, {
+                        expected: turnLease?.generation,
+                        current: this.turnCoordinator.current(session).generation,
+                    });
                     return;
                 }
                 if (!commitDecision.ok) {
@@ -3559,7 +3648,10 @@ export class ResponseEngine {
                         }
                         this.agentDelegationRegistry?.revokeTask(session.id, taskId);
                         logger.info(`[ResponseEngine] Stale turn stopped before protocol rejection: session=${session.id} task=${taskId}`);
-                        await this.publishTriggerExecutionInterrupted(message, adapter, envelope, 'stale_generation', taskCausation);
+                        await this.publishTriggerExecutionInterrupted(message, adapter, envelope, 'stale_generation', taskCausation, {
+                            expected: turnLease?.generation,
+                            current: this.turnCoordinator.current(session).generation,
+                        });
                         return;
                     }
                     const reason = streamResult.protocolIncompleteReason || commitDecision.reason;
@@ -3761,7 +3853,10 @@ export class ResponseEngine {
                 if (!commitDecision.ok) {
                     logger.info(`[ResponseEngine] Turn commit rejected after flush: session=${session.id} task=${taskId} reason=${commitDecision.reason}`);
                     if (commitDecision.reason === 'stale_generation') {
-                        await this.publishTriggerExecutionInterrupted(message, adapter, envelope, commitDecision.reason, taskCausation);
+                        await this.publishTriggerExecutionInterrupted(message, adapter, envelope, commitDecision.reason, taskCausation, {
+                            expected: turnLease?.generation,
+                            current: this.turnCoordinator.current(session).generation,
+                        });
                     }
                     else {
                         this.publishTriggerExecutionFailure(message, `turn_commit_rejected:${commitDecision.reason}`, {
@@ -3848,6 +3943,8 @@ export class ResponseEngine {
                     this.eventBus.publish({
                         type: 'task:error',
                         sessionId: session.id,
+                        taskId,
+                        generation: turnLease?.generation,
                         error: errorSummary,
                         errorType,
                         agentName: agentNameForStats,
@@ -3965,12 +4062,18 @@ export class ResponseEngine {
                             cache_read_tokens: sum.cache_read_tokens,
                             cache_creation_tokens: sum.cache_creation_tokens,
                             // 顶层 cost_usd/cost_cny 保持向后兼容 = 网关实际价
-                            cost_usd: sum.cost_gateway_usd,
-                            cost_cny: sum.cost_gateway_cny,
+                            cost_usd: roundCostForOutput(sum.cost_gateway_usd),
+                            cost_cny: roundCostForOutput(sum.cost_gateway_cny),
                             call_count: sum.calls,
                             cost: {
-                                official: { usd: sum.cost_official_usd, cny: sum.cost_official_cny },
-                                gateway: { usd: sum.cost_gateway_usd, cny: sum.cost_gateway_cny },
+                                official: {
+                                    usd: roundCostForOutput(sum.cost_official_usd),
+                                    cny: roundCostForOutput(sum.cost_official_cny),
+                                },
+                                gateway: {
+                                    usd: roundCostForOutput(sum.cost_gateway_usd),
+                                    cny: roundCostForOutput(sum.cost_gateway_cny),
+                                },
                             },
                         };
                     }
@@ -3988,10 +4091,13 @@ export class ResponseEngine {
                     }
                     else {
                         // cost 同时给原价(official)与网关实际价(gateway)；顶层 cost_usd/cost_cny 保持向后兼容 = 网关价。
-                        const gatewayUsd = turnCost.gateway?.usd ?? turnCost.official?.usd ?? 0;
-                        const gatewayCny = turnCost.gateway?.cny ?? turnCost.official?.cny ?? 0;
+                        const gatewayUsd = roundCostForOutput(turnCost.gateway?.usd ?? turnCost.official?.usd ?? 0);
+                        const gatewayCny = roundCostForOutput(turnCost.gateway?.cny ?? turnCost.official?.cny ?? 0);
                         const turnCostBlock = {
-                            official: { usd: turnCost.official?.usd ?? 0, cny: turnCost.official?.cny ?? 0 },
+                            official: {
+                                usd: roundCostForOutput(turnCost.official?.usd ?? 0),
+                                cny: roundCostForOutput(turnCost.official?.cny ?? 0),
+                            },
                             gateway: { usd: gatewayUsd, cny: gatewayCny },
                         };
                         // 最后一次访问：本轮可能有多次大模型调用（numTurns>1），整轮的 turnCostBlock 不等于
@@ -4006,10 +4112,13 @@ export class ResponseEngine {
                                     model: lastModel, turns: 1,
                                 });
                                 const lp = resolvePrices(resolveRoot(), lastEvent, agent.getGatewayPricing?.());
-                                const lpGwUsd = lp.gateway?.usd ?? lp.official?.usd ?? 0;
-                                const lpGwCny = lp.gateway?.cny ?? lp.official?.cny ?? 0;
+                                const lpGwUsd = roundCostForOutput(lp.gateway?.usd ?? lp.official?.usd ?? 0);
+                                const lpGwCny = roundCostForOutput(lp.gateway?.cny ?? lp.official?.cny ?? 0);
                                 lastModelCall = { ...lastModelCall, cost: {
-                                        official: { usd: lp.official?.usd ?? 0, cny: lp.official?.cny ?? 0 },
+                                        official: {
+                                            usd: roundCostForOutput(lp.official?.usd ?? 0),
+                                            cny: roundCostForOutput(lp.official?.cny ?? 0),
+                                        },
                                         gateway: { usd: lpGwUsd, cny: lpGwCny },
                                     } };
                             }
@@ -4057,6 +4166,8 @@ export class ResponseEngine {
                 this.eventBus.publish({
                     type: 'task:completed',
                     sessionId: session.id,
+                    taskId,
+                    generation: turnLease.generation,
                     channel: message.channel,
                     channelId: message.channelId,
                     terminalReason: streamResult.terminalReason,
@@ -4188,6 +4299,8 @@ export class ResponseEngine {
                 this.eventBus.publish({
                     type: 'task:error',
                     sessionId: session.id,
+                    taskId,
+                    generation: turnLease?.generation,
                     error: errorMsg,
                     errorType,
                     agentName: agentNameForStats,
@@ -4208,14 +4321,15 @@ export class ResponseEngine {
             }
             // 发送用户友好的错误消息
             // 用户主动中断（新消息打断 或 /stop 命令）时静默，不发送错误提示
-            // processEventStream 已通过 renderer 发过错误时也跳过
+            // 普通渠道可跳过 renderer 已发送的重复错误；daemon trigger 仍必须
+            // 发送结构化终态，否则 DaemonChannel 会一直等到 watchdog。
             const retryExhaustedCount = getRetryExhaustedCount(error);
             const modelFallbackExhaustedMessage = getModelFallbackExhaustedMessage(error);
             const retryInputAlreadySubmitted = wasRetryInputAlreadySubmitted(error);
             if (isUserInterrupt) {
                 logger.info(`[ResponseEngine] User interrupt by new_message, skip sending error message`);
             }
-            else if (error?._errorAlreadySent && !retryExhaustedCount && !isTimeout && !isTotalExecutionTimeout) {
+            else if (error?._errorAlreadySent && !daemonTrigger && !retryExhaustedCount && !isTimeout && !isTotalExecutionTimeout) {
                 logger.info(`[ResponseEngine] Error already sent via renderer, skip sending duplicate message`);
             }
             else {
@@ -4633,11 +4747,17 @@ export class ResponseEngine {
                 if (event.type === 'complete') {
                     event = normalizeCompleteAgentEvent(event);
                 }
-                if (event.type === 'tool_result' && event.isError && !event.errorCode) {
-                    event = {
-                        ...event,
-                        errorCode: classifyToolErrorCode({ error: event.error, result: event.result }),
-                    };
+                if (event.type === 'tool_result' && event.isError) {
+                    const classificationMissing = !normalizeToolErrorCode(event.errorCode)
+                        && event.decisionSource !== 'policy'
+                        && event.decisionSource !== 'approval';
+                    const errorCode = classifyToolErrorCode({
+                        errorCode: event.errorCode,
+                        decisionSource: event.decisionSource,
+                    });
+                    if (event.errorCode !== errorCode) {
+                        event = { ...event, errorCode, ...(classificationMissing ? { classificationMissing: true } : {}) };
+                    }
                 }
                 // 每收到事件重置空闲超时
                 const toolName = event.type === 'tool_use' ? event.name : undefined;
@@ -4740,7 +4860,13 @@ export class ResponseEngine {
                     eventDetail = ` tool=${event.name}${desc ? ` desc="${desc}"` : ''}`;
                 }
                 else if (event.type === 'tool_result') {
-                    eventDetail = ` tool=${event.name} ok=${!event.isError}`;
+                    const decision = event.decision ?? (event.isError ? 'error' : 'allow');
+                    const executed = event.executed ?? !event.isError;
+                    const executionState = event.executionState
+                        ?? (decision === 'deny' ? 'blocked' : event.isError ? 'failed' : 'completed');
+                    eventDetail = ` tool=${event.name} ok=${decision === 'allow' && executed}`
+                        + ` decision=${decision} executed=${executed} executionState=${executionState}`
+                        + (event.policyCode ? ` policy=${event.policyCode}` : '');
                 }
                 const frameworkEvents = new Set(['session_id', 'state_changed', 'status']);
                 if (frameworkEvents.has(event.type)) {
@@ -5006,6 +5132,16 @@ export class ResponseEngine {
                                 toolName: event.name,
                                 toolInput: event.input || {},
                                 injectToModel: (text) => { agent.injectUserMessage?.(session.id, text); },
+                                recordReminder: reminder => {
+                                    this.eventBus.publish({
+                                        type: 'runner:proactive-reminder',
+                                        sessionId: session.id,
+                                        ...reminder,
+                                        injection: 'requested',
+                                        timestamp: Date.now(),
+                                        causation,
+                                    });
+                                },
                                 getQueueLength: () => this.messageQueue?.getQueueLength(session.id) ?? 0,
                                 isSendCommand: (toolName, toolInput) => isExactEvolcoreSendCommandForSession(toolName, toolInput, session.channelId, proactiveState?.chatType === 'group' ? 'group' : 'private', proactiveSelfAid || session.selfAID),
                                 logger,
@@ -5031,6 +5167,7 @@ export class ResponseEngine {
                             sessionId: session.id,
                             callId: event.callId,
                             correlationId: event.correlationId,
+                            requestId: event.requestId,
                         });
                         this.eventBus.publish({
                             type: 'tool:result',
@@ -5039,16 +5176,29 @@ export class ResponseEngine {
                             sessionId: session.id,
                             toolName: event.name,
                             isError: event.isError,
-                            ...(event.isError ? { errorCode: classifyToolErrorCode({ errorCode: event.errorCode, error: event.error, result: event.result }) } : {}),
+                            ...(event.isError ? {
+                                errorCode: classifyToolErrorCode({
+                                    errorCode: event.errorCode,
+                                    decisionSource: event.decisionSource,
+                                }),
+                                ...((event.classificationMissing || (!normalizeToolErrorCode(event.errorCode)
+                                    && event.decisionSource !== 'policy'
+                                    && event.decisionSource !== 'approval')) ? { classificationMissing: true } : {}),
+                            } : {}),
                             ...(event.callId ? { callId: event.callId } : {}),
                             ...(event.correlationId || event.callId ? { correlationId: event.correlationId ?? event.callId } : {}),
+                            ...(event.requestId ? { requestId: event.requestId } : {}),
                             agentName: lifecycleAgentName,
                             agentAid: lifecycleAgentAid ?? 'unknown',
                             permissionMode: lifecyclePermissionMode,
-                            decision: event.isError ? 'error' : 'allow',
-                            decisionSource: 'runner',
-                            executed: true,
-                            executionState: event.isError ? 'failed' : 'completed',
+                            decision: event.decision ?? (event.isError ? 'error' : 'allow'),
+                            decisionSource: event.decisionSource ?? 'runner',
+                            ...(event.policyCode ? { policyCode: event.policyCode } : {}),
+                            ...(event.reason ? { reason: event.reason } : {}),
+                            executed: event.executed ?? (event.decision === 'deny' ? false : true),
+                            executionState: event.executionState ?? (event.decision === 'deny'
+                                ? 'blocked'
+                                : event.isError ? 'failed' : 'completed'),
                             timestamp: Date.now(),
                             causation,
                         });
@@ -5095,7 +5245,23 @@ export class ResponseEngine {
                                 result: event.result,
                                 isError: event.isError,
                                 error: event.error,
+                                decision: event.decision,
+                                decisionSource: event.decisionSource,
+                                executed: event.executed,
+                                executionState: event.executionState,
+                                policyCode: event.policyCode,
+                                requestId: event.requestId,
                                 injectToModel: (text) => { agent.injectUserMessage?.(session.id, text); },
+                                recordReminder: reminder => {
+                                    this.eventBus.publish({
+                                        type: 'runner:proactive-reminder',
+                                        sessionId: session.id,
+                                        ...reminder,
+                                        injection: 'requested',
+                                        timestamp: Date.now(),
+                                        causation,
+                                    });
+                                },
                                 getQueueLength: () => this.messageQueue?.getQueueLength(session.id) ?? 0,
                                 isSendCommand: (toolName, toolInput) => isExactEvolcoreSendCommandForSession(toolName, toolInput, session.channelId, proactiveState?.chatType === 'group' ? 'group' : 'private', proactiveSelfAid || session.selfAID),
                                 logger,
@@ -5123,8 +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,
@@ -5134,14 +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），
@@ -5160,34 +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()) {