mocode-ai 1.6.8 → 1.6.9

package/dist/llm/providers/vertex.js

@@ -254,8 +254,12 @@ export function buildGeminiRequest(messages, tools, runtime, overrides) {
     const geminiTools = encodeGeminiTools(tools);
     const effort = effectiveReasoningEffort(overrides?.reasoningEffort);
     const budget = thinkingBudget(effort);
+    // Gemini 的思考 token 计入 maxOutputTokens：不给思考留余量时，思考先吃光预算，
+    // finishReason=MAX_TOKENS 且正文为空/截断（3.8-flash 实测：maxOutputTokens=100 → 正文空，
+    // =2000 → 92 thoughts + "OK"）。auto 档动态思考上限按 flash 系 24576 保守留量。
+    const thinkingAllowance = budget ?? (effort === 'auto' ? 24576 : 0);
     const generationConfig = {
-        maxOutputTokens: runtimeConfig.maxTokens ?? 8192,
+        maxOutputTokens: (runtimeConfig.maxTokens ?? 8192) + thinkingAllowance,
         temperature: 0.2,
         ...(budget === null ? {} : { thinkingConfig: { thinkingBudget: budget } }),
     };
@@ -329,6 +333,7 @@ export async function vertexChatOnce(messages, handlers, signal, tools, runtime,
     let usage;
     const toolCalls = [];
     let reportedTool = false;
+    let finishReason = '';
     const live = { cjk: 0, other: 0 };
     const countLive = (text) => {
         for (const ch of text) {
@@ -355,6 +360,8 @@ export async function vertexChatOnce(messages, handlers, signal, tools, runtime,
             usage = usageFromGemini(usageMetadata);
         const candidates = Array.isArray(event.candidates) ? event.candidates : [];
         const candidate = isRecord(candidates[0]) ? candidates[0] : undefined;
+        if (typeof candidate?.finishReason === 'string')
+            finishReason = candidate.finishReason;
         const candidateContent = isRecord(candidate?.content) ? candidate.content : undefined;
         const parts = Array.isArray(candidateContent?.parts) ? candidateContent.parts : [];
         for (const rawPart of parts) {
@@ -393,5 +400,12 @@ export async function vertexChatOnce(messages, handlers, signal, tools, runtime,
             cachedTokens: usage.cachedTokens,
         });
     }
+    // MAX_TOKENS 截断曾静默返回：模型没发完工具调用就结束 turn（「没干完活就结束」）。
+    // 显式标注让用户与下一轮模型都看到截断事实；正常 STOP/SAFETY 等不加。
+    if (finishReason === 'MAX_TOKENS' && (content || toolCalls.length > 0)) {
+        const marker = '\n\n[mocode] Gemini finishReason=MAX_TOKENS：输出在 token 上限处被截断（思考也计入 maxOutputTokens）。可调大 MAX_TOKENS 或 /effort 降低思考强度。';
+        content += marker;
+        handlers.onText?.(marker);
+    }
     return { content: content || null, toolCalls, usage };
 }

package/dist/repl/commands/model-actions.js

@@ -34,6 +34,14 @@ export function applyPresetConfig(ctx, p) {
         CONTEXT_WINDOW_TOKENS: String(p.contextWindow),
         ANTHROPIC_PROMPT_CACHE: p.anthropicPromptCache ? 'true' : 'false',
     });
+    // 关键不变量：应用预设即激活——必须同步 .active 指针。否则重启时旧激活预设的优先级
+    // 高于刚写入的 config 文件，表现为「切了模型、重启又回到原来的」（实测的丢配置 bug）。
+    try {
+        setActivePresetName(p.name);
+    }
+    catch {
+        /* 指针写失败不阻断切换 */
+    }
     reconfigureClient();
     ctx.refreshStatusBase(ctx.history);
     layout.clearContent();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mocode-ai",
-  "version": "1.6.8",
+  "version": "1.6.9",
   "description": "终端编码 agent:LLM + tool-call 循环 + 流式输出(含思考)+ 27 个工具,接任意 OpenAI 兼容后端。",
   "type": "module",
   "bin": {