[Bug]: Using version 1.10.5, the issue frequently occurs where the JSON format returned by the LLM is incorrect, causing midscene to report errors and fail.

Version

System:
    OS: Windows 11 10.0.26200
    CPU: (8) x64 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz
    Memory: 2.10 GB / 15.71 GB
  Browsers:
    Chrome: 150.0.7871.116
    Edge: Chromium (150.0.4078.65)
    Internet Explorer: 11.0.26100.8115
  npmPackages:
    @midscene/cli: ^1.10.5 => 1.10.5
    @midscene/web: ^1.10.5 => 1.10.5

Details

使用1.10.5版本，频繁出现LLM返回的json格式不对导致midscene报错失败的问题，之前的版本没有出现过，是否是新版本出现的问题？ json的{} 变成了[]，LLM格式问题是否应该加校验或格式清洗手段

Reproduce link

无

Reproduce Steps

LLM信息
MIDSCENE_MODEL_NAME=qwen3.7-plus
MIDSCENE_MODEL_FAMILY=qwen3
发送的是点击事件
ai-call.log
[2026-07-20T19:21:57.585+08:00] adapter chat completion params: {"config":{"temperature":0,"enable_thinking":false}}
[2026-07-20T19:21:57.585+08:00] sending request to qwen3.7-plus
[2026-07-20T19:22:00.080+08:00] response reasoning content:
[2026-07-20T19:22:00.080+08:00] response content: ```json
[
"bbox": [345, 529, 391, 565],
"errors": []
}

[2026-07-20T19:22:00.081+08:00] adapter chat completion params: {"config":{"temperature":0,"enable_thinking":false}}
[2026-07-20T19:22:00.082+08:00] sending request to qwen3.7-plus
[2026-07-20T19:22:00.245+08:00] response reasoning content: 
[2026-07-20T19:22:00.245+08:00] response content: ```json
[
  "bbox": [345, 529, 391, 565],
  "errors": []
}

ServiceError: failed to locate element:
failed to parse LLM response into JSON. Error - Error: Unexpected character "`" at position 59. Response -

[
 "bbox": [345, 529, 391, 565],
 "errors": []
}

at Service.locate (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\service\index.ts:181:13)
at processTicksAndRejections (node:internal/process/task_queues:104:5)
at Object.executor (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\agent\task-builder.ts:551:28)
at TaskRunner.flush (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\task-runner.ts:335:25)
at TaskExecutor.runPlans (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\agent\tasks.ts:349:20)
at PuppeteerPageAgent.callActionInActionSpace (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\agent\agent.ts:697:24)
at D:\project_code\midscene-monitor\node_modules\@midscene\core\src\agent\agent.ts:719:7
at PuppeteerPageAgent.aiTap (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\agent\agent.ts:718:5)
at ScriptPlayer.playFlowItem (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\yaml\player.ts:585:7)
at ScriptPlayer.playTask (D:\project_code\midscene-monitor\node_modules\@midscene\core\src\yaml\player.ts:308:9)

Start at src/service/index.ts:181, where the stack trace reports the LLM response parsing failure, and inspect the response handling used by the locate flow. Reproduce with the logged fenced response if possible, then determine the expected object shape and how malformed JSON should be handled. Done means the reported response no longer causes an unhandled parsing failure and the behavior is covered by a regression test.