evolcore 0.0.24 → 0.0.26

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

@@ -1275,6 +1275,8 @@ export class MessageQueue {
             return false;
         if (first.projectPath !== next.projectPath || first.agentName !== next.agentName)
             return false;
+        // Results from the same parent handoff are still mergeable. Their
+        // resultRefs are retained and consumed individually after one turn.
         const a = first.message.handoffDelivery?.originAuthorization;
         const b = next.message.handoffDelivery?.originAuthorization;
         if (!a || !b)
@@ -1323,6 +1325,10 @@ export class MessageQueue {
         let hasMessagesAfterRestartResume = false;
         const originHandoffIds = [];
         const seenOriginHandoffIds = new Set();
+        const originResultIds = [];
+        const seenOriginResultIds = new Set();
+        const originResultRefs = [];
+        const seenOriginResultRefs = new Set();
         const allOriginHandoffs = items.every(item => this.isOriginHandoff(item.message));
         for (const item of items) {
             const m = item.message;
@@ -1347,6 +1353,24 @@ export class MessageQueue {
                         originHandoffIds.push(id);
                     }
                 }
+                const refs = m.handoffDelivery.resultRefs?.length
+                    ? m.handoffDelivery.resultRefs
+                    : (m.handoffDelivery.resultIds?.length
+                        ? m.handoffDelivery.resultIds.map(resultId => ({ handoffId: m.handoffDelivery.handoffId, resultId }))
+                        : (m.handoffDelivery.resultId
+                            ? [{ handoffId: m.handoffDelivery.handoffId, resultId: m.handoffDelivery.resultId }]
+                            : []));
+                for (const ref of refs) {
+                    const refKey = `${ref.handoffId}\u0000${ref.resultId}`;
+                    if (!seenOriginResultRefs.has(refKey)) {
+                        seenOriginResultRefs.add(refKey);
+                        originResultRefs.push(ref);
+                    }
+                    if (!seenOriginResultIds.has(ref.resultId)) {
+                        seenOriginResultIds.add(ref.resultId);
+                        originResultIds.push(ref.resultId);
+                    }
+                }
             }
             for (const id of [...(m.sourceMessageIds ?? []), ...(m.messageId ? [m.messageId] : [])]) {
                 if (!seenSourceMessageIds.has(id)) {
@@ -1418,6 +1442,8 @@ export class MessageQueue {
                     ...last.message.handoffDelivery,
                     handoffId: originHandoffIds[0],
                     handoffIds: originHandoffIds,
+                    ...(originResultIds.length > 0 ? { resultIds: originResultIds } : {}),
+                    ...(originResultRefs.length > 0 ? { resultRefs: originResultRefs } : {}),
                 }
                 : last.message.handoffDelivery,
         };

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

@@ -101,6 +101,7 @@ export function deliveryTargetFromSession(session) {
 export function replyContextFromSession(session, context) {
     return {
         ...(session.metadata?.replyContext ?? {}),
+        ...(session.threadId ? { threadId: session.threadId } : {}),
         ...(context ?? {}),
         // Session chatType/groupId is the persisted routing authority. The nested
         // reply context can outlive a route correction, so only an explicit

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

@@ -2,7 +2,7 @@ import path from 'path';
 import fs from 'fs';
 import os from 'os';
 import crypto from 'crypto';
-import { BaseagentRunnerUnavailableError, hasCompact, autoCompactWindowForContextWindow, autoCompactWindowForModel, isClaudeContextUsageModel, numericToken } from '../../agents/runner-types.js';
+import { BaseagentRunnerUnavailableError, hasCompact, autoCompactWindowForModel, isClaudeContextUsageModel, numericToken } from '../../agents/runner-types.js';
 import { buildTaskRuntimeEnv, ensureRuntimeLockDir, ensureSessionRuntimeDir } from '../../cli/task-context.js';
 import { IMRenderer } from './im-renderer.js';
 import { createSendFileMarkerPattern } from './file-markers.js';
@@ -11,7 +11,8 @@ import { StreamIdleMonitor } from './stream-idle-monitor.js';
 import { logger } from '../../utils/logger.js';
 import { isHostChinese } from '../../utils/locale.js';
 import { getErrorMessage, classifyError, ErrorType, ERROR_PREFIX, isInfraError, isAbortTerminalReason, prefixErrorType, isRetryableError, isContextTooLongText } from '../../utils/error-utils.js';
-import { isExactEvolcoreSendCommandForSession } from '../permission/ec-command-parser.js';
+import { readEvolcoreRuntimeFlavor, tagErrorTextWithEvolcoreVersion } from '../../utils/evolcore-version.js';
+import { describeProactiveSendCommandFailure, isExactEvolcoreSendCommandForSession } from '../permission/ec-command-parser.js';
 import { summarizeToolInput } from '../../utils/tool-summary.js';
 import { agentDir as resolveAgentDir, agentMdPath, getPackageRoot, resolveRoot, resolvePaths } from '../../paths.js';
 import { renderKitSections } from '../../eck/kit-renderer.js';
@@ -485,9 +486,6 @@ export async function resolveGroupRulesForTurn(opts, sync = syncGroupRulesContex
     }
     return sync({ selfAid: opts.selfAid, groupId: opts.groupId, channel: opts.channel });
 }
-function autoCompactTokensFromMaxTokens(maxTokens) {
-    return autoCompactWindowForContextWindow(maxTokens);
-}
 /**
  * 构造 OutboundEnvelope —— 出站三件套的信封部分。
  *
@@ -915,6 +913,20 @@ export class ResponseEngine {
             causation: runtime?.causation,
         });
     }
+    async recoverAuthorizedHandoffResult(params, selfAid) {
+        if (!this.handoffRuntime) {
+            return { ok: false, code: 'HANDOFF_NOT_FOUND', error: 'handoff runtime not configured' };
+        }
+        const runtime = this.activeTaskRuntimeContexts.get(params.sessionId);
+        return this.handoffRuntime.recoverReturnHandoff({
+            selfAid,
+            recoverySessionId: params.sessionId,
+            handoffId: params.handoffId,
+            content: params.content,
+            actorId: params.actorId,
+            causation: runtime?.causation,
+        });
+    }
     agentRegistry;
     setAgentRegistry(registry) {
         this.agentRegistry = registry;
@@ -2132,6 +2144,7 @@ export class ResponseEngine {
             modeConfig: resolvedMode.context.modeConfig,
             state: modeState,
             isSendCommand: (toolName, toolInput) => isExactEvolcoreSendCommandForSession(toolName, toolInput, session.channelId, chatType, selfAid),
+            describeSendCommandFailure: describeProactiveSendCommandFailure,
             logger,
         } : null;
         if (resolvedMode?.mode.beforeProcess && modeProcessCtx) {
@@ -2764,6 +2777,7 @@ export class ResponseEngine {
             let previousFallbackState;
             let agentModel;
             let v2HandoffIds = [];
+            let v2HandoffResultRefs = [];
             let v2HandoffDirection;
             let handoffPromptRendered = false;
             let runtimeEnv;
@@ -2998,6 +3012,12 @@ export class ResponseEngine {
                         bootstrapAgentMdPath: agentMdPath(selfAid),
                     };
                 }
+                if (selfAid && peerRole === 'owner' && effectivePermissionMode !== 'readonly') {
+                    modelOverride = {
+                        ...(modelOverride || {}),
+                        personalMdPath: path.join(resolveRoot(), 'agents', selfAid, 'personal', 'personal.md'),
+                    };
+                }
                 if (normalizedBaseagent.canonical === 'claude') {
                     modelOverride = {
                         ...(modelOverride || {}),
@@ -3092,7 +3112,7 @@ export class ResponseEngine {
                         KITS_FRAGMENTS: path.join(pkgRoot, 'kits', 'templates', 'system-fragments'),
                         KITS_MESSAGE_FRAGMENTS: path.join(pkgRoot, 'kits', 'templates', 'message-fragments'),
                         // evolcore 运行模式：dev=源码仓库 | install=全局安装包
-                        evolcoreMode: fs.existsSync(path.join(pkgRoot, 'src', 'index.ts')) ? 'dev' : 'install',
+                        evolcoreMode: readEvolcoreRuntimeFlavor(),
                         // 路径变量(用于 manifest 路径展开,resolvePath 用 ctx.vars 取真值)
                         PERSONAL_DIR: selfAid ? path.join(resolveRoot(), 'agents', selfAid, 'personal') : undefined,
                         RELATIONS_DIR: selfAid ? path.join(resolveRoot(), 'agents', selfAid, 'relations') : undefined,
@@ -3235,6 +3255,7 @@ export class ResponseEngine {
                                 v2HandoffIds = Array.from(new Set(items.flatMap(item => (item.handoff?.handoffIds?.length
                                     ? item.handoff.handoffIds
                                     : (item.handoff?.handoffId ? [item.handoff.handoffId] : [])))));
+                                v2HandoffResultRefs = items.flatMap(item => item.handoff?.resultRefs ?? []);
                                 v2HandoffDirection = message.handoffDelivery.direction;
                                 return items;
                             }
@@ -3308,11 +3329,18 @@ export class ResponseEngine {
                     return;
                 }
                 if (v2HandoffIds.length > 0 && handoffPromptRendered && this.handoffRuntime && selfAid) {
-                    for (const handoffId of v2HandoffIds) {
-                        if (v2HandoffDirection === 'origin')
-                            this.handoffRuntime.completeOriginContext(selfAid, handoffId);
-                        else if (v2HandoffDirection === 'target')
-                            this.handoffRuntime.completeTargetContext(selfAid, handoffId);
+                    if (v2HandoffDirection === 'origin' && v2HandoffResultRefs.length > 0) {
+                        for (const ref of v2HandoffResultRefs) {
+                            this.handoffRuntime.completeOriginContext(selfAid, ref.handoffId, ref.resultId);
+                        }
+                    }
+                    else {
+                        for (const handoffId of v2HandoffIds) {
+                            if (v2HandoffDirection === 'origin')
+                                this.handoffRuntime.completeOriginContext(selfAid, handoffId);
+                            else if (v2HandoffDirection === 'target')
+                                this.handoffRuntime.completeTargetContext(selfAid, handoffId);
+                        }
                     }
                 }
                 // The runner may already own a stronger per-session directory (Codex
@@ -4446,7 +4474,9 @@ export class ResponseEngine {
                         kind: 'activity.batch',
                         items: [{
                                 kind: 'notice',
-                                text: userMessage,
+                                // activity.batch 载荷混合正常内容与错误，顶层出站包裹不整体打标；
+                                // 这里的 notice 是 proactive 下用户能看到的错误副本，单独打标。
+                                text: tagErrorTextWithEvolcoreVersion(userMessage),
                                 severity: 'warn',
                                 subtype: 'task-error',
                             }],
@@ -4562,7 +4592,7 @@ export class ResponseEngine {
             const shouldClose = !writerDb;
             try {
                 const row = db.prepare(`SELECT model, input_tokens, cache_creation_tokens, cache_read_tokens,
-                  context_tokens, max_tokens, auto_compact_tokens
+                  context_tokens
              FROM model_calls
             WHERE session_id = ?
               AND agent_session_id = ?
@@ -4584,16 +4614,11 @@ export class ResponseEngine {
                 }
                 if (totalTokens <= 0)
                     return undefined;
-                const recordedAutoCompactTokens = row.auto_compact_tokens ?? undefined;
-                const inferredAutoCompactTokens = autoCompactTokensFromMaxTokens(row.max_tokens ?? undefined);
-                const candidateAutoCompactTokens = recordedAutoCompactTokens && recordedAutoCompactTokens > 0
-                    ? recordedAutoCompactTokens
-                    : inferredAutoCompactTokens ?? autoCompactWindowForModel(model);
-                // Historic rows may contain the old fixed 350k threshold. Never let a
-                // persisted value exceed the context window that accompanied it.
-                const autoCompactTokens = inferredAutoCompactTokens
-                    ? Math.min(candidateAutoCompactTokens, inferredAutoCompactTokens)
-                    : candidateAutoCompactTokens;
+                // EvolCore 的自动压缩阈值固定为 650k，与模型上下文窗口无关：
+                // 大窗口模型（1M）到 650k 由 EvolCore 压缩；小窗口模型（200k/128k）
+                // 到不了该阈值，交给 base agent 自身的压缩机制处理。
+                // 历史行里可能存着旧的 350k 阈值，这里一律忽略，只认当前常量。
+                const autoCompactTokens = autoCompactWindowForModel(model);
                 return { totalTokens, autoCompactTokens };
             }
             finally {

package/dist/core/model/catalog-service.js

@@ -3,11 +3,21 @@ import fs from 'node:fs';
 import path from 'node:path';
 import { appendRequestPath, appendV1RequestPath } from '../../agents/request-identity.js';
 import { atomicReadJson, atomicWriteJson } from '../../utils/atomic-write.js';
-export const CATALOG_TTL_MS = 5 * 60_000;
-export const CATALOG_STALE_MS = 24 * 60 * 60_000;
+import { parseModelId } from './context-window-suffix.js';
+/** 单页请求的超时（含响应体读取）。多页目录里每页各自享有这个预算。 */
+export const CATALOG_PAGE_TIMEOUT_MS = 4_000;
+/** 整条路由（含全部分页）的总期限，兜住慢网关无限翻页。 */
+export const CATALOG_ROUTE_TIMEOUT_MS = 15_000;
+/**
+ * 持久化快照的格式版本。**目录规范化或能力过滤规则变化时必须递增**，
+ * 否则旧规则写下的持久化快照与新快照无法区分，会继续被服务。
+ * v2：保留上下文窗口后缀（`claude-opus-5[1m]` 不再被折叠成基础 ID）。
+ * v3：保留 Codex 上游声明的 max effort，重新获取旧快照中过滤掉的能力。
+ */
+export const CATALOG_SNAPSHOT_VERSION = 3;
 export const TRANSPORT_EFFORTS = {
     claude: ['low', 'medium', 'high', 'xhigh', 'max'],
-    codex: ['none', 'minimal', 'low', 'medium', 'high', 'xhigh'],
+    codex: ['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'],
     ecagent: ['minimal', 'low', 'medium', 'high', 'xhigh', 'max'],
     gemini: [],
 };
@@ -22,7 +32,7 @@ export const catalogHash = (value) => createHash('sha256').update(stable(value))
 export function catalogContextKey(context) {
     const headers = Object.fromEntries(Object.entries(context.headers ?? {}).map(([k, v]) => [k.toLowerCase(), v]));
     return catalogHash({ scope: context.scope, baseagent: context.baseagent, provider: context.provider,
-        baseUrl: context.baseUrl?.replace(/\/+$/, ''), apiKey: context.apiKey,
+        baseUrl: context.baseUrl?.replace(/\/+$/, ''), apiKey: context.apiKey, authType: context.authType,
         headers, queryParams: context.queryParams ?? {}, configUnavailable: context.configUnavailable });
 }
 const strings = (value) => Array.isArray(value)
@@ -40,37 +50,45 @@ export function parseModelList(json, _url) {
     const out = new Map();
     const items = rows(json);
     for (const item of items) {
-        const rawId = typeof item === 'string' ? item : item?.id ?? item?.slug ?? item?.name ?? item?.model;
-        if (typeof rawId !== 'string' || !rawId.trim())
-            continue;
-        // 规范化模型 ID：去掉 Claude Code SDK 的上下文窗口后缀 [1m]/[200k]
-        // 这些后缀只在传给 claude baseagent 时由 applyContextWindow 添加
-        const id = rawId.trim().replace(/\[(?:1m|200k)\]$/i, '');
-        if (out.has(id))
-            continue;
-        const rawLevels = item?.reasoningLevels ?? item?.supportedReasoningEfforts ?? item?.supported_reasoning_levels;
-        const levels = Array.isArray(rawLevels) ? strings(rawLevels.map(v => typeof v === 'string' ? v : v?.reasoningEffort ?? v?.effort)) : undefined;
-        const owner = item?.owned_by ?? item?.owner ?? item?.provider;
-        out.set(id, {
-            id, verification: 'listed',
-            ...(typeof owner === 'string' ? { owned_by: owner } : {}),
-            ...(typeof item?.created === 'number' ? { created: item.created } : {}),
-            ...(typeof (item?.displayName ?? item?.display_name) === 'string' ? { displayName: item.displayName ?? item.display_name } : {}),
-            ...(levels ? { reasoningLevels: levels } : {}),
-            ...(typeof (item?.defaultReasoningEffort ?? item?.default_reasoning_level) === 'string' ? { defaultReasoningLevel: item.defaultReasoningEffort ?? item.default_reasoning_level } : {}),
-            ...(strings(item?.protocols ?? item?.supported_protocols) ? { protocols: strings(item.protocols ?? item.supported_protocols) } : {}),
-            ...(typeof item?.capabilities?.tool_calling === 'boolean' ? { supportsTools: item.capabilities.tool_calling } : {}),
-            ...(typeof (item?.context_window ?? item?.contextWindow) === 'number' ? { contextWindow: item.context_window ?? item.contextWindow } : {}),
-            ...(typeof item?.max_output_tokens === 'number' ? { maxOutputTokens: item.max_output_tokens } : {}),
-        });
+        // 单个畸形条目只跳过自己：一条读不懂的记录不该让整份目录退化到 builtin。
+        try {
+            const rawId = typeof item === 'string' ? item : item?.id ?? item?.slug ?? item?.name ?? item?.model;
+            if (typeof rawId !== 'string' || !rawId.trim())
+                continue;
+            // 上下文窗口后缀 [1m]/[200k] 是模型身份的一部分：网关把不同窗口作为独立条目
+            // 下发，两个变体必须各自成为一条可选条目。仅规范化后缀大小写以便去重。
+            const parsed = parseModelId(rawId.trim());
+            const id = `${parsed.base}${parsed.suffix ?? ''}`;
+            if (out.has(id))
+                continue;
+            const rawLevels = item?.reasoningLevels ?? item?.supportedReasoningEfforts ?? item?.supported_reasoning_levels;
+            const levels = Array.isArray(rawLevels) ? strings(rawLevels.map(v => typeof v === 'string' ? v : v?.reasoningEffort ?? v?.effort)) : undefined;
+            const owner = item?.owned_by ?? item?.owner ?? item?.provider;
+            out.set(id, {
+                id, verification: 'listed',
+                ...(typeof owner === 'string' ? { owned_by: owner } : {}),
+                ...(typeof item?.created === 'number' ? { created: item.created } : {}),
+                ...(typeof (item?.displayName ?? item?.display_name) === 'string' ? { displayName: item.displayName ?? item.display_name } : {}),
+                ...(levels ? { reasoningLevels: levels } : {}),
+                ...(typeof (item?.defaultReasoningEffort ?? item?.default_reasoning_level) === 'string' ? { defaultReasoningLevel: item.defaultReasoningEffort ?? item.default_reasoning_level } : {}),
+                ...(strings(item?.protocols ?? item?.supported_protocols) ? { protocols: strings(item.protocols ?? item.supported_protocols) } : {}),
+                ...(typeof item?.capabilities?.tool_calling === 'boolean' ? { supportsTools: item.capabilities.tool_calling } : {}),
+                // 网关显式给出的 context_window 优先；缺失时用后缀声明的窗口兜底。
+                ...(typeof (item?.context_window ?? item?.contextWindow) === 'number' ? { contextWindow: item.context_window ?? item.contextWindow }
+                    : parsed.contextWindow !== undefined ? { contextWindow: parsed.contextWindow } : {}),
+                ...(typeof item?.max_output_tokens === 'number' ? { maxOutputTokens: item.max_output_tokens } : {}),
+            });
+        }
+        catch { /* Skip the offending entry; the rest of the page stays usable. */ }
     }
+    // 全部条目都无法解析仍视为不可信响应：避免把一个格式不符的 200 当成空目录。
     if (items.length && !out.size)
         throw new Error('Model list contains no valid IDs');
     return [...out.values()];
 }
 const BUILTIN_IDS = {
     codex: ['gpt-5.5', 'gpt-5.4', 'gpt-5.4-mini', 'gpt-5.3-codex', 'gpt-5.2'],
-    claude: ['claude-opus-4-7', 'claude-opus-4-6', 'claude-sonnet-4-6', 'claude-haiku-4-5-20251001', 'deepseek-v4-pro', 'deepseek-v4-flash', 'kimi-k2.6', 'kimi-k2.5', 'glm-5.1', 'glm-5', 'glm-4.7', 'MiniMax-M2.7'],
+    claude: ['claude-opus-5', 'claude-sonnet-5', 'claude-opus-4-8', 'claude-opus-4-7', 'claude-opus-4-6', 'claude-sonnet-4-6', 'claude-haiku-4-5-20251001', 'deepseek-v4-pro', 'deepseek-v4-flash', 'kimi-k2.6', 'kimi-k2.5', 'glm-5.1', 'glm-5', 'glm-4.7', 'MiniMax-M2.7'],
     ecagent: ['claude-sonnet-4-6'],
     gemini: ['gemini-2.5-pro', 'gemini-2.5-flash', 'gemini-2.5-flash-lite'],
 };
@@ -89,9 +107,14 @@ export function deriveAliases(models, baseagent) {
         if (!match)
             continue;
         const [, family, major, minor = '0'] = match;
+        // 同版本的多个窗口变体：取窗口最大的一条，别名才不会随网关下发顺序漂移。
+        const window = parseModelId(id).contextWindow ?? 0;
         const current = best[family];
-        if (!current || +major > current.major || (+major === current.major && +minor > current.minor))
-            best[family] = { id, major: +major, minor: +minor };
+        const better = !current || +major > current.major
+            || (+major === current.major && +minor > current.minor)
+            || (+major === current.major && +minor === current.minor && window > current.window);
+        if (better)
+            best[family] = { id, major: +major, minor: +minor, window };
     }
     return Object.fromEntries(Object.entries(best).map(([family, info]) => [family, info.id]));
 }
@@ -133,12 +156,17 @@ async function gatewayModels(context, fetcher) {
             const models = [];
             const seen = new Set();
             let next = url;
-            // Bound the whole route, including response bodies and pagination.
-            const signal = AbortSignal.timeout(4000);
+            // 每页各自计时，避免多页目录被一个覆盖全程的超时提前掐断；
+            // 同时用总期限兜住整条路由，防止慢网关无限翻页。
+            const deadline = Date.now() + CATALOG_ROUTE_TIMEOUT_MS;
             for (let page = 0; page < 100; page++) {
                 if (seen.has(next))
                     throw new GatewayError();
                 seen.add(next);
+                const remaining = deadline - Date.now();
+                if (remaining <= 0)
+                    throw new GatewayError();
+                const signal = AbortSignal.timeout(Math.min(CATALOG_PAGE_TIMEOUT_MS, remaining));
                 const headers = new Headers(context.headers);
                 if (context.apiKey && !headers.has('authorization')) {
                     headers.set('authorization', `Bearer ${context.apiKey}`);
@@ -176,7 +204,7 @@ async function gatewayModels(context, fetcher) {
 /** One instance per daemon. Offline clients use the same persistence format. */
 export class ModelCatalogService {
     options;
-    cache = new Map();
+    latest = new Map();
     successes = new Map();
     inflight = new Map();
     constructor(options = {}) {
@@ -192,41 +220,43 @@ export class ModelCatalogService {
             try {
                 const file = this.file(context);
                 const saved = file ? atomicReadJson(file) : undefined;
-                if (saved?.version === 1 && saved.snapshot && Array.isArray(saved.snapshot.models) && ['gateway', 'native'].includes(saved.snapshot.source))
+                // 版本不匹配的快照按冷启动处理：其 ID 可能出自已废弃的规范化规则。
+                if (saved?.version === CATALOG_SNAPSHOT_VERSION && saved.snapshot && Array.isArray(saved.snapshot.models) && ['gateway', 'native'].includes(saved.snapshot.source))
                     snapshot = saved.snapshot;
             }
             catch { /* A missing/corrupt cache is a cold load. */ }
             if (snapshot)
                 this.successes.set(key, snapshot);
         }
-        return snapshot && this.now() - Date.parse(snapshot.fetchedAt) <= CATALOG_STALE_MS ? snapshot : undefined;
+        return snapshot;
     }
     peek(context) {
-        return structuredClone(this.cache.get(this.key(context))?.snapshot ?? this.previous(context) ?? this.make(context, 'builtin', 'fallback', this.legacyBuiltins(context), [{ code: 'BUILTIN', message: '内置候选，未经当前连接验证' }]));
+        return structuredClone(this.latest.get(this.key(context)) ?? this.previous(context) ?? this.make(context, 'builtin', 'fallback', this.legacyBuiltins(context), [{ code: 'BUILTIN', message: '内置候选，未经当前连接验证' }]));
     }
     async get(context, refresh = false) {
         const key = this.key(context);
-        const cached = this.cache.get(key);
-        if (!refresh && cached && cached.retryAt > this.now())
-            return structuredClone(cached.snapshot);
-        const previous = this.previous(context);
-        if (!refresh && previous && this.now() - Date.parse(previous.fetchedAt) < CATALOG_TTL_MS && !cached) {
-            this.cache.set(key, { snapshot: previous, retryAt: Date.parse(previous.fetchedAt) + CATALOG_TTL_MS });
-            return structuredClone(previous);
-        }
+        const previous = this.latest.get(key) ?? this.previous(context);
         let pending = this.inflight.get(key);
         if (!pending) {
             pending = this.load(context, previous).then(snapshot => {
-                this.cache.set(key, { snapshot, retryAt: this.now() + (snapshot.freshness === 'fresh' ? CATALOG_TTL_MS : 30_000) });
+                // Keep any usable first result (including builtin/native fallback), but
+                // never let a failed refresh replace the last usable snapshot.
+                if (snapshot.freshness !== 'stale')
+                    this.latest.set(key, snapshot);
                 return snapshot;
             }).finally(() => this.inflight.delete(key));
             this.inflight.set(key, pending);
         }
-        if (!refresh && previous)
-            return structuredClone({ ...previous, freshness: 'stale', diagnostics: [...(cached?.snapshot.diagnostics ?? []), { code: 'STALE_CACHE', message: '显示同一连接的最近成功目录，正在刷新' }] });
+        if (!refresh && previous) {
+            // Serve the last successful catalog immediately on every ordinary read,
+            // while each read also asks the shared background job to refresh it.
+            // Explicit refresh callers still await that same in-flight request.
+            void pending.catch(() => undefined);
+            return structuredClone(previous);
+        }
         return structuredClone(await pending);
     }
-    clear() { this.cache.clear(); this.successes.clear(); }
+    clear() { this.latest.clear(); this.successes.clear(); }
     make(context, source, freshness, models, diagnostics) {
         const entries = adapt(models, context.baseagent);
         const fetchedAt = new Date(this.now()).toISOString();
@@ -241,7 +271,7 @@ export class ModelCatalogService {
             const file = this.file(context);
             if (file) {
                 fs.mkdirSync(path.dirname(file), { recursive: true, mode: 0o700 });
-                atomicWriteJson(file, { version: 1, snapshot });
+                atomicWriteJson(file, { version: CATALOG_SNAPSHOT_VERSION, snapshot });
             }
         }
         catch { /* Persistence failure does not discard a live result. */ }

package/dist/core/model/context-window-suffix.js

@@ -0,0 +1,33 @@
+/**
+ * Claude Code SDK 的上下文窗口后缀（如 `claude-opus-5[1m]`）。
+ *
+ * 网关会把同一模型的不同窗口作为独立条目下发，所以后缀是**模型身份的一部分**，
+ * 目录与选择界面必须原样保留——两个变体各成一条可选条目。
+ * 只有按基础 ID 归档的场景（查价、计费、能力参数表）才剥离后缀。
+ */
+/** 后缀 token → 该后缀声明的上下文窗口。 */
+export const CONTEXT_WINDOW_SUFFIX_TOKENS = {
+    '1m': 1_000_000,
+    '200k': 200_000,
+};
+const SUFFIX_RE = /\[(1m|200k)\]$/i;
+/** 拆分模型 ID 的基础部分与上下文窗口后缀。 */
+export function parseModelId(modelId) {
+    const match = SUFFIX_RE.exec(modelId);
+    if (!match)
+        return { base: modelId };
+    const token = match[1].toLowerCase();
+    return {
+        base: modelId.slice(0, match.index),
+        suffix: `[${token}]`,
+        contextWindow: CONTEXT_WINDOW_SUFFIX_TOKENS[token],
+    };
+}
+/** 去掉上下文窗口后缀。仅用于查价/计费/能力表这类按基础 ID 归档的场景。 */
+export function stripContextWindowSuffix(modelId) {
+    return modelId.replace(SUFFIX_RE, '');
+}
+/** 该 ID 是否已带网关显式给出的上下文窗口后缀。 */
+export function hasContextWindowSuffix(modelId) {
+    return SUFFIX_RE.test(modelId);
+}

package/dist/core/model/model-catalog.js

@@ -88,7 +88,7 @@ export function resolveCatalogContext(self, ba) {
             : baseagent === 'claude' ? resolveAnthropicConfig(cfg, block.claude)
                 : baseagent === 'ecagent' ? resolveEcagentConfig(cfg, block.ecagent) : undefined;
         if (resolved) {
-            Object.assign(context, { baseUrl: resolved.baseUrl, apiKey: resolved.apiKey, queryParams: resolved.queryParams,
+            Object.assign(context, { baseUrl: resolved.baseUrl, apiKey: resolved.apiKey, authType: 'authType' in resolved ? resolved.authType : undefined, queryParams: resolved.queryParams,
                 provider: 'provider' in resolved ? resolved.provider : undefined,
                 headers: buildModelRequestHeaders({ baseagent: baseagent, baseUrl: resolved.baseUrl, agentAid: self, configuredHeaders: resolved.headers }) });
         }

package/dist/core/model/model-permission.js

@@ -1,10 +1,19 @@
 import { getFieldPermission } from '../../config/roles.js';
+import { stripContextWindowSuffix } from './context-window-suffix.js';
 export function isModelAllowedByPatterns(model, allowedModels) {
     if (allowedModels.includes('*'))
         return true;
-    return allowedModels.some(pattern => pattern.endsWith('*')
-        ? model.startsWith(pattern.slice(0, -1))
-        : model === pattern);
+    // 窗口变体是同一模型的不同上下文窗口：白名单写基础 ID（`claude-opus-5`）即同时
+    // 授权其窗口变体（`claude-opus-5[1m]`）。反向不成立——白名单显式写了某个变体时，
+    // 只授权那一个变体，其它变体和裸 ID 仍被拒绝。
+    const candidates = new Set([model, stripContextWindowSuffix(model)]);
+    return allowedModels.some(pattern => {
+        if (pattern.endsWith('*')) {
+            const prefix = pattern.slice(0, -1);
+            return [...candidates].some(candidate => candidate.startsWith(prefix));
+        }
+        return candidates.has(pattern);
+    });
 }
 function normalizeBaseagent(baseagent) {
     const normalized = typeof baseagent === 'string' ? baseagent.trim() : '';