fix(ai-chat): explain json failures and widen judge

The interjection judge capped output at 300 tokens, which reasoning
models can spend on thinking before writing any JSON, so judging kept
failing with a bare "no valid JSON" message. Raise the cap to 1024 and
make JSON failures report finish_reason, a short output sample and the
number of tries, so truncation and empty replies show in the log.

Author: xzdmycbx

src/plugins/ai-chat/llm-client.test.ts

@@ -200,3 +200,14 @@ describe('isInputError', () => {
     expect(isInputError(new TypeError("Header 'Authorization' has invalid value"))).toBe(false);
   });
 });
+
+describe('jsonCompletion 的失败信息', () => {
+  test('被 maxTokens 截断时，错误里写明结束原因和输出片段', async () => {
+    scriptFetch([[200, { choices: [{ message: { content: '{"action":"re' }, finish_reason: 'length' }] }]]);
+    const err = await jsonCompletion({ ...baseReq, purpose: 'chat', repairAttempts: 0, validate: () => null }).catch((e) => e);
+    expect(err).toBeInstanceOf(JsonOutputError);
+    expect((err as Error).message).toContain('finish_reason=length');
+    expect((err as Error).message).toContain('截断');
+    expect((err as Error).message).toContain('{"action":"re');
+  });
+});

src/plugins/ai-chat/llm-client.ts

@@ -331,6 +331,9 @@ export async function jsonCompletion<T>(req: JsonRequest<T>): Promise<T> {
   const messages: ChatMessage[] = [...base.messages];
   const key = `${base.target.apiUrl}|${base.target.model}`;
 
+  // 最后一次的输出与结束原因：失败时写进错误，日志里能看出是被截断、空回复还是答非所问
+  let lastText = '';
+  let lastFinish = '';
   for (let round = 0; round <= repairAttempts; round++) {
     const useJsonMode = !jsonModeUnsupported.has(key);
     let data: ChatCompletionResponse;
@@ -355,6 +358,8 @@ export async function jsonCompletion<T>(req: JsonRequest<T>): Promise<T> {
     const parsed = parseJsonLoose(text);
     const value = parsed === undefined ? null : validate(parsed);
     if (value !== null) return value;
+    lastText = text;
+    lastFinish = data.choices?.[0]?.finish_reason || '';
 
     if (round < repairAttempts) {
       logger.warn('AI', `${logPrefix(base.purpose)} 输出不是合格的 JSON，要求重新生成`);
@@ -367,7 +372,10 @@ export async function jsonCompletion<T>(req: JsonRequest<T>): Promise<T> {
     }
   }
 
-  throw new JsonOutputError('模型多次未返回符合要求的 JSON');
+  const tries = repairAttempts + 1;
+  const hint = lastFinish === 'length' ? '：输出被 maxTokens 截断（推理模型的思考也占 token）' : '';
+  const sample = lastText ? `，输出：${lastText.replace(/\s+/g, ' ').slice(0, 120)}` : '，输出为空';
+  throw new JsonOutputError(`模型${tries > 1 ? ` ${tries} 次都` : ''}没有返回符合要求的 JSON（finish_reason=${lastFinish || '无'}${hint}${sample}）`);
 }
 
 /** 测试专用：清空"不支持 response_format"的记忆 */

src/plugins/ai-chat/proactive.ts

@@ -369,7 +369,8 @@ export async function judge(persona: PersonaCard, scan: TriggerScan, now: number
       { role: 'system', content: p.system },
       { role: 'user', content: p.user },
     ],
-    maxTokens: 300,
+    // 输出只是一小段 JSON，但推理模型的思考也算在 maxTokens 里：给小了会在写出 JSON 之前被截断
+    maxTokens: 1024,
     temperature: 0.5,
     repairAttempts: 0,
     validate: (v) => validateJudge(v, p.refs, getAiConfig().proactive.allowReactOnly),