Repository navigation
Releases
v0.6.0
设置页的开关(模型分组、推理档位共享、输入框积分、上下文选档)即时生效并保存,重启 DSH 后保持。积分在启动后自动统计,并每 5 分钟刷新一次(creditsIntervalMs)。写操作失败时单独提示,不再显示成「读取状态失败」。「全部检测」可以停止;登录成功后提供「完成」和「再添加一个」。
修复:导入授权文件时的路径穿越(uid 白名单);自动续期请求加了超时,并有看门狗兜底;续期写回被拒时不再记为成功;卸载时可能出现的 unhandledRejection;启动中途出错后可自动重试;「N 个模型有可选窗口档」提示不显示;删掉了不正确的「需重启 DSH」提示;关掉推理档位共享不再影响上下文选档。
界面:设置区移到模型列表之前,布尔设置改成开关;账号行的积分列对齐,状态标签只保留一个,窄屏下改为两行;标签与复选框不再使用主题强调色;模型列表去掉无意义的圆点,选档下拉放在行内。
Settings-page switches now apply at once and persist across restarts. Credits are counted at startup and refreshed every 5 minutes (creditsIntervalMs). Write failures get their own notice. "Detect all" can be stopped; the login panel offers Done / Add another.
Fixes: path traversal via account uid on import; keepalive request timeout plus a sweep watchdog; keepalive no longer counts a rejected write-back as success; an unhandledRejection on abort; startup errors now retry; the window-tier count hint; a stale "restart DSH" hint; turning off reasoning sharing no longer disables context tiers.
UI: settings above the model list, switches for boolean settings, aligned credit column, one health tag per account, two-line rows on narrow widths, neutral tags and checkboxes, inline tier selects, no meaningless dots.
v0.5.2
修复:模型行上的窗口下拉和「全部最高/最低」按钮点了没反应(客户端漏注册动作名)。文案改成平实写法;折叠的模型列表标题会提示有几个模型可选窗口。
Fix: tier selects actually fire; plainer copy; tier count in the collapsed header.
v0.5.1
上下文选档改为设置项:设置页里给每个模型选窗口档(也有「全部最高/全部最低/恢复默认」),选择器里每个模型仍然只有一个条目一个名字(0.5.0 的拆分方案作废)。另外插件设置页注册进了 DSH 设置侧边栏。
Context tiers as a per-model setting (picker stays one name per model); settings sidebar entry.
v0.5.0
上下文选档:目录声明了多档窗口的模型(如 GPT-6.1-Sol 的 272K/600K/1M)在选择器里拆成多个条目(· 600K / · 1M),每个模型各自选窗口。合成条目经一个只绑回环的小代理把请求里的模型 id 换回真实 id 再逐字节转发,鉴权与 SSE 原样透传。设置页可开关。
Per-model context window tiers in the picker, via split entries and a loopback model-id rewriting proxy.
v0.4.9
「端点不校验参数」的模型(Kimi-K3、GLM-5.2 等)实测档位是真生效的(kimi-k3 实测 low→max 思考量 133→400 递增),现在按惯例直接提供 low / medium / high / max 四档,选择器里就能选,不用检测;标签统一为深色款。
Fixed-default reasoning models now get conventional low/medium/high/max levels (verified effective by measurement); tags all use the dark style.
v0.4.8
始终思考但档位固定的模型(如 GLM-5.2/5.1、MiniMax-M3)现在会显示「思考·medium」标签,不再像「没有思考」。
Always-on reasoning models with a fixed level now show a tag instead of nothing.
v0.4.7
修复:检测区的折叠收不回来(完成后只自动展开一次,之后随意收);模型行现在会把检测出来的档位也贴上(和目录声明的同款外观);「全部检测」期间的事件广播合并成一次,页面不再被带着重排。
Collapsible probe section fixed, detected levels shown on model rows, event broadcast debounced.
v0.4.6
已检测的模型不再占用候选列表(收进默认折叠的「已检测(N)」组,可展开重新检测);删掉危险的「清除已探测结果」按钮(它会一键清掉所有账号的全部记录,误点即全毁);检测区默认折叠、检测时自动展开。
Detected models move to a collapsed group; the dangerous clear-all button is removed; the probe section collapses by default.
v0.4.5
修复:设置页的推理档位检测区误显示「当前没有可检测的模型」——过滤器把面板自己的数据也清了。面板现在走未过滤通道,候选列表恢复正常。
Fix: the probe panel wrongly showed an empty candidate list; the settings surface now bypasses the composer filter.
v0.4.4
「推理等级」按钮不再出现在输入框(检测只在设置页进行);修复等待动画图标撑满整行的尺寸问题。
The composer's reasoning-level button is gone for good — detection lives in settings only; fixed the oversized spinner icon.