v0.14.1 想得再深、也得出片 / Think deep, still deliver
「想得再深、也得出片」——修复多家平台「深度思考」模型触发的「找爆点失败:LLM 未返回内容」(#8)。 / Think deep, still deliver.
根因 / Root cause
- 各平台的默认/热门模型如今多是「深度思考(推理)」型,思考过程的 token 也计入输出预算——长转写稿一思考就把 4000 token 预算全烧光,正文(content)为空,于是报「LLM 未返回内容」;换平台也躲不开,因为大家默认给的都是思考模型
- Most platforms now default to reasoning models whose thinking counts against the output budget; on a long transcript they burn all 4,000 tokens thinking and return empty content — switching platforms doesn't help because everyone defaults to a thinker
空响应三层自救 / Three-layer empty-response salvage
- 大预算自动重试——检测到「只有思考、没有正文」(finish=length)时,自动换 16000 token 预算重试一次,多数场景直接救活;重试自身的报错不覆盖「空响应」这个更准的诊断
- reasoning 字段兜底——个别网关把正文错放进 reasoning 字段(模型正常收尾但 content 为空),直接取回交给下游解析
- 分因报错——仍然失败时按证据写明原因和下一步:思考模型烧完预算(换非思考版,通常带 instruct/chat 字样)/内容被平台安全审查拦截(换供应商或素材)/服务端偶发空响应(点重试)
- Auto-retry with a 16k budget when the response is thinking-only; recover content misplaced into the
reasoningfield by some gateways; and if still empty, the error names the actual cause with a concrete next step (switch to a non-thinking variant / content filter hit / transient blip)
完整功能清单见 README · Full feature list in the README
下载 / Download: macOS 拿 .dmg,Windows 拿 .exe(安装版)或 .zip(绿色版),Linux 拿 .AppImage(实验性,#7)。