feat: v0.7.0 四阶段全量迭代 — 修复面收口 · 安全纵深 · 架构还债 · 能力演进

P1 修复面收口: v0.6.3 截断自愈推全量(Anthropic/Ollama/非流式/引擎兜底); SSE 上游错误帧检测进重试通道;
clearMessages 摘要游标根治; truncateResult 内联图片白名单统一; 前端四 bug(确认弹窗锁死/MemoryViewer/
Virtuoso Footer/abort 尾部过滤) + reasoning 缓冲跨迭代污染; 托盘通知过滤与新建会话死链接线

P2 安全纵深: MCP 审批闭环(ConfirmationHook×PolicyEngine 联动+重名拒注册); SSRF 收敛 ssrf-guard 共享模块
(web_fetch 双通道校验+重定向终态复检); Electron 加固(preload CJS 化→sandbox:true/CSP/权限白名单/will-navigate);
run_command cmd.exe 白名单通道元字符守门; diff_viewer 10MB 预检; Anthropic thinking 预算下限; Agnes 思考显式关闭

P3 架构还债: OpenAICompatibleAdapter 中间基类收敛四家样板; 错误分类单轨化(删 mapError/getFetchSignal,
超时显式 ETIMEDOUT); PRAGMA user_version 迁移版本化; 死代码清理专项(cn.ts/SHORTCUTS/ContextMenu 分支/
getWindowState/modifiedArgs/sandbox 空壳); i18next 引入; a11y 第一轮; SearXNG 页批量草稿模型统一

P4 能力演进: Ollama pull 可取消/capabilities 探测/num_ctx 实测缓存; UpdateService feed 比对式自动更新
(app:updateCheck IPC + StatusBar 入口); MiMo providerOptions(web_search 服务端工具/strict JSON);
web_fetch extract_mode=markdown(turndown); network.proxyUrl 全局代理(Chromium sessions+undici dispatcher)

测试: 264 → 507 用例(Electron ABI 全绿零跳过), 覆盖引擎压缩管线/重试竞速/MEMORY.md 闸门/file_editor 五操作/
filesystem 七工具实体夹具/git 真实仓库/SSE 错误帧/全线截断自愈/Provider 请求形态矩阵/SSRF 表测/钩子分级矩阵/
OutputValidator 全量/SLO 指标/MCP 安全纯函数/task_manager 链路/渲染层纯域/i18n 桥契约

electron/harness/adapters/anthropic.adapter.ts

@@ -16,7 +16,8 @@
  * @see https://docs.anthropic.com/en/api/messages
  */
 
-import { BaseAdapter } from './base-adapter';
+import { BaseAdapter, ContentFilterError } from './base-adapter';
+import { truncatedArgumentsPayload } from './shared/sse-stream';
 import log from 'electron-log';
 import { nanoid } from 'nanoid';
 import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
@@ -119,6 +120,12 @@ export class AnthropicAdapter extends BaseAdapter {
     // 工具调用缓冲：content block index → { id, name, argsBuffer }
     const toolBlocks = new Map<number, { id: string; name: string; argsBuffer: string }>();
 
+    // v0.6.4 竞态修复: message_start 捕获的 input_tokens 改为本次调用的局部闭包变量。
+    // 原实现放在实例字段（this.lastInputTokens）—— fallback adapter 是跨引擎共享的
+    // 单例（agent-engine-manager 把同一实例注入所有引擎），故障转移后多个并发会话
+    // 共用该 Anthropic 实例时 input_tokens 会互相串号。局部化后天然隔离。
+    let messageStartInputTokens = 0;
+
     const base = () => ({
       requestId: request.meta.requestId,
       sessionId: request.meta.sessionId,
@@ -127,6 +134,37 @@ export class AnthropicAdapter extends BaseAdapter {
       timestamp: Date.now(),
     });
 
+    /**
+     * v0.6.4 错误事件单轨化: Anthropic `error` SSE 事件不再以普通 ERROR 流事件转发
+     * （引擎对 ERROR 事件的旧处理是 throw 普通 Error，最终落入 UNKNOWN 且完全绕过
+     * chatStreamWithRetry 的重试/故障转移）。改为抛出携带归一化 status 的异常，
+     * 与 HTTP 层 throwHttpError 同轨：overloaded/rate_limit 走重试、authentication/
+     * invalid_request 不重试并可触发 fallback。
+     */
+    const anthropicErrorCodeToStatus = (code: string): number => {
+      switch (code) {
+        case 'overloaded_error':
+          return 529;
+        case 'rate_limit_error':
+          return 429;
+        case 'api_error':
+          return 500;
+        case 'timeout_error':
+          return 504;
+        case 'authentication_error':
+          return 401;
+        case 'permission_error':
+          return 403;
+        case 'not_found_error':
+          return 404;
+        case 'request_too_large':
+        case 'invalid_request_error':
+          return 400;
+        default:
+          return 500;
+      }
+    };
+
     const processEvent = (name: string, data: Record<string, unknown>): MetonaStreamEvent[] => {
       const events: MetonaStreamEvent[] = [];
       switch (name) {
@@ -173,8 +211,17 @@ export class AnthropicAdapter extends BaseAdapter {
             let args: Record<string, unknown> = {};
             try {
               args = block.argsBuffer ? JSON.parse(block.argsBuffer) : {};
-            } catch {
-              args = {};
+            } catch (err) {
+              // v0.6.4 缺口 A 修复: content_block_stop 时 argsBuffer 解析失败（流截断致
+              // JSON 半截）—— 原实现静默降级 args={}，与 v0.6.3 已修复的 OpenAI 共享层
+              // 行为完全相同：工具以"缺少必要参数"泛化失败，模型无从得知发生了截断，
+              // 长文件写入场景直接导致"空回复 → 会话无声终止"。现统一转为
+              // _truncatedArguments 错误参数触发模型自愈（与共享层同源、同文案契约）。
+              const sample = block.argsBuffer.slice(-120);
+              log.warn(
+                `[Anthropic] Tool call args truncated at content_block_stop (unparseable JSON, ${(err as Error).message}). Tail: ...${sample}`,
+              );
+              args = truncatedArgumentsPayload((err as Error).message, sample);
             }
             events.push({
               type: MetonaStreamEventType.TOOL_CALL_COMPLETE,
@@ -199,10 +246,14 @@ export class AnthropicAdapter extends BaseAdapter {
               type: MetonaStreamEventType.USAGE,
               ...base(),
               usage: {
-                inputTokens: (this.lastInputTokens as number) ?? 0,
+                inputTokens: messageStartInputTokens,
                 outputTokens: (usage.output_tokens as number) ?? 0,
                 totalTokens:
-                  ((this.lastInputTokens as number) ?? 0) + ((usage.output_tokens as number) ?? 0),
+                  messageStartInputTokens + ((usage.output_tokens as number) ?? 0),
+                // v0.6.4: 补采 Anthropic 自己的缓存字段（其他 provider 均已采集，
+                // cache_read/creation_input_tokens 与 output_tokens 同在 usage 内）
+                cacheHitTokens: (usage.cache_read_input_tokens as number) ?? undefined,
+                cacheMissTokens: (usage.cache_creation_input_tokens as number) ?? undefined,
               },
             });
           }
@@ -215,16 +266,18 @@ export class AnthropicAdapter extends BaseAdapter {
         }
         case 'error': {
           const err = data.error as Record<string, unknown> | undefined;
-          events.push({
-            type: MetonaStreamEventType.ERROR,
-            ...base(),
-            error: {
-              code: 'unknown' as never,
-              message: (err?.message as string) ?? 'Anthropic stream error',
-              retryable: false,
-            },
-          });
-          break;
+          const code = (err?.type as string) ?? 'api_error';
+          const message = (err?.message as string) ?? 'Anthropic stream error';
+          const status = anthropicErrorCodeToStatus(code);
+          log.warn(
+            `[Anthropic] Upstream error event: ${code} (normalized status=${status}) — throwing for retry/failover handling`,
+          );
+          if (code === 'content_filter_error') {
+            throw new ContentFilterError(message, 'Anthropic SSE error event');
+          }
+          const throwable = new Error(`anthropic_stream_error (${code}): ${message}`);
+          (throwable as Error & { status: number }).status = status;
+          throw throwable;
         }
       }
       return events;
@@ -256,13 +309,15 @@ export class AnthropicAdapter extends BaseAdapter {
           if (eventName === 'message_start') {
             const msg = data.message as Record<string, unknown> | undefined;
             const usage = msg?.usage as Record<string, unknown> | undefined;
-            this.lastInputTokens = (usage?.input_tokens as number) ?? 0;
+            messageStartInputTokens = (usage?.input_tokens as number) ?? 0;
             continue;
           }
           for (const ev of processEvent(eventName, data)) {
             yield ev;
           }
         } catch (parseErr) {
+          // ContentFilterError / 带 status 的上游错误由 processEvent 抛出，需原样透传
+          if (parseErr instanceof Error && parseErr.name !== 'SyntaxError') throw parseErr;
           log.warn(
             `[Anthropic] Failed to parse SSE line: ${(parseErr as Error).message}`,
             trimmed.slice(0, 200),
@@ -271,15 +326,48 @@ export class AnthropicAdapter extends BaseAdapter {
       }
     }
 
-    // 流中断（连接断开等）补发 DONE，防止 Agent Loop 挂起（与 Ollama 行为一致）
+    // v0.6.4 缺口 B 修复: 流中断时不再让缓冲中的 tool_use 整体蒸发。
+    // 原实现在 content_block_start 与 content_block_stop 之间断连时，toolBlocks 里
+    // 未完成的 block 既不产生 TOOL_CALL_COMPLETE、也不 flush —— 引擎看到"零工具调用
+    // + 零文本"→ 误判 COMPLETED 空回复 → 会话无声停止（正是 v0.6.3 宣称根治、但在
+    // Anthropic 流上仍然存活的场景）。现于补发 DONE 之前，将所有未完成块按截断契约
+    // 转为 _truncatedArguments 自愈 tool call（解析成功的则正常产出）。
     if (!streamEndedNormally) {
+      const unfinished = [...toolBlocks.entries()];
+      if (unfinished.length > 0) {
+        log.warn(
+          `[Anthropic] Stream ended without message_stop with ${unfinished.length} unfinished tool block(s) — flushing as truncated/self-healing tool calls`,
+        );
+        for (const [, block] of unfinished) {
+          let args: Record<string, unknown> = {};
+          try {
+            args = block.argsBuffer ? JSON.parse(block.argsBuffer) : {};
+          } catch (err) {
+            args = truncatedArgumentsPayload(
+              (err as Error).message,
+              block.argsBuffer.slice(-120),
+            );
+          }
+          yield {
+            type: MetonaStreamEventType.TOOL_CALL_COMPLETE,
+            ...base(),
+            toolCall: {
+              id: block.id,
+              name: block.name,
+              args,
+              iteration: request.meta.iteration,
+              timestamp: Date.now(),
+            },
+          };
+        }
+      } else {
+        log.warn('[Anthropic] Stream ended without message_stop (connection likely dropped)');
+      }
+      toolBlocks.clear();
       yield { type: MetonaStreamEventType.DONE, ...base() };
     }
   }
 
-  /** message_start 捕获的 input_tokens（供 message_delta 汇总 usage） */
-  private lastInputTokens = 0;
-
   // ===== 模型与上下文窗口 =====
 
   override async listModels(): Promise<MetonaModelInfo[]> {
@@ -311,6 +399,8 @@ export class AnthropicAdapter extends BaseAdapter {
     request: MetonaRequest,
     stream: boolean,
   ): Promise<Record<string, unknown>> {
+    const thinkingRequested = Boolean(request.params.thinkingEnabled);
+
     // System Prompt 拼接（Anthropic 使用顶层 system 字段）
     const system = [
       request.systemPrompt.roleDefinition,
@@ -405,9 +495,17 @@ export class AnthropicAdapter extends BaseAdapter {
     const anthropicMaxOutput =
       AnthropicAdapter.MODEL_INFO[this.config.defaultModel]?.maxOutputTokens ?? 64_000;
 
+    // v0.6.4 边界加固: thinking 开启时保证 max_tokens ≥ 2048 —— 协议要求
+    // budget_tokens >= 1024 且 < max_tokens。原实现当用户配置极小 maxTokens
+    // （如 1500）时 Math.floor(1500/2)=750 < 1024 直接 API 400。
+    const requestedMaxTokens = request.params.maxTokens ?? 8192;
+    const maxTokensForRequest = thinkingRequested
+      ? Math.max(2048, Math.min(requestedMaxTokens, anthropicMaxOutput))
+      : Math.min(requestedMaxTokens, anthropicMaxOutput);
+
     const body: Record<string, unknown> = {
       model: this.config.defaultModel,
-      max_tokens: Math.min(request.params.maxTokens ?? 8192, anthropicMaxOutput),
+      max_tokens: maxTokensForRequest,
       system,
       messages: merged,
       stream,
@@ -423,17 +521,15 @@ export class AnthropicAdapter extends BaseAdapter {
     }
 
     // Thinking 模式：budget_tokens（必须小于 max_tokens，此处钳制到一半）
-    if (request.params.thinkingEnabled) {
+    if (thinkingRequested) {
       const budgetMap: Record<string, number> = {
         low: 1024,
         medium: 4096,
         high: 16384,
         max: 32768,
       };
-      const budget = Math.min(
-        budgetMap[request.params.thinkingEffort ?? 'high'] ?? 16384,
-        Math.floor((body.max_tokens as number) / 2),
-      );
+      const effortBudget = budgetMap[request.params.thinkingEffort ?? 'high'] ?? 16384;
+      const budget = Math.min(effortBudget, Math.floor(maxTokensForRequest / 2));
       body.thinking = { type: 'enabled', budget_tokens: budget };
     } else {
       body.temperature = request.params.temperature;
@@ -485,9 +581,13 @@ export class AnthropicAdapter extends BaseAdapter {
 
     for (const block of contentBlocks) {
       if (block.type === 'text') text += (block.text as string) ?? '';
-      else if (block.type === 'thinking')
-        reasoningContent = (block.thinking as string) ?? undefined;
-      else if (block.type === 'tool_use') {
+      else if (block.type === 'thinking') {
+        // v0.6.4 修复: 多个 thinking 块应为累加（原实现后者覆盖前者，长推理链丢内容）
+        const thinking = (block.thinking as string) ?? '';
+        if (thinking) {
+          reasoningContent = reasoningContent ? `${reasoningContent}\n\n${thinking}` : thinking;
+        }
+      } else if (block.type === 'tool_use') {
         let args: Record<string, unknown> = {};
         const rawInput = block.input;
         if (rawInput && typeof rawInput === 'object') args = rawInput as Record<string, unknown>;
@@ -503,12 +603,16 @@ export class AnthropicAdapter extends BaseAdapter {
 
     const usage = (data.usage as Record<string, number>) ?? {};
     const stopReason = (data.stop_reason as string) ?? 'end_turn';
+    // v0.6.4: refusal / content_filter 不再折叠为 STOP —— 语义丢失会让上层把
+    // "被拒绝的回答"当正常回复展示；统一映射为 CONTENT_FILTERED 走友好提示链路
     const finishReason: MetonaFinishReason =
       stopReason === 'tool_use'
         ? MetonaFinishReason.TOOL_CALLS
         : stopReason === 'max_tokens'
           ? MetonaFinishReason.LENGTH
-          : MetonaFinishReason.STOP;
+          : stopReason === 'refusal' || stopReason === 'content_filter'
+            ? MetonaFinishReason.CONTENT_FILTER
+            : MetonaFinishReason.STOP;
 
     return {
       meta: {
@@ -525,6 +629,9 @@ export class AnthropicAdapter extends BaseAdapter {
         inputTokens: usage.input_tokens ?? 0,
         outputTokens: usage.output_tokens ?? 0,
         totalTokens: (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0),
+        // v0.6.4: 补采缓存字段（与非流式调用方对齐其他 provider 的口径）
+        cacheHitTokens: usage.cache_read_input_tokens,
+        cacheMissTokens: usage.cache_creation_input_tokens,
       },
       finishReason,
     };