Skip to content

v1.1.4

Choose a tag to compare

View all tags
@qingyu321 qingyu321 released this 12 Aug 10:38
· 73 commits to main since this release
v1.1.4
ded2428

v1.1.4 — 滚动追踪全面修复 + 轮次历史展开 + 工具期绿点

新功能

  • 轮次历史展开预览 — 点击右侧轮次序号向左展开面板，查看该轮次的完整用户提问（当前轮优先倒序展示），可滚动浏览全部历史
  • 桌宠 token 合计 — token 面板显示该后端全部任务的合计用量（输入 / 输出 / 缓存），多会话并行时一眼看到总消耗
  • 桌宠持久化汇报 — 任务完成 / 出错汇报在气泡中持久显示，直到新任务完成替换或手动隐藏，后台任务跑完不再错过

修复

  • 对话滚动追踪全面修复 — 提交后等待首个 token 与工具调用阶段不再误判为「用户滚走」，自动跟随不再中断；思考阶段不再上下抖动，随时可以拉到底；「最新」按钮一键到底并自动收尾
  • 滚动跟随不再被抢占 — 排队消息 / 自动重试 / 自动压缩触发的新轮次不会把正在读历史的你拉回底部；滚动条拖动、触控板惯性、键盘滚动与输入区滚轮都能正确暂停自动跟随；切换会话标签恢复原阅读位置；图片加载、卡片展开收起后自动补滚到底
  • 工具执行期绿点保持 — 侧栏「活跃中」绿点在工具执行期间不再熄灭，模型以工具调用结尾的一轮也会保持绿点，直到该轮真正结束
  • 联网搜索能力检测式兜底 — 启动时探测主 API 是否原生支持搜索：支持则直连，不支持才自动切换兜底端点执行搜索，失败自动重试、状态自愈
  • 桌宠修复 — 默认关闭双保险；右键菜单双向钳制不再被裁剪；气泡与徽章不再重叠
  • 安全与稳定性批次 — 命令注入防护（默认应用打开、Git 调用加固）、zip 解压路径穿越加固、API 密钥明文导出确认改为原生对话框、预发布版本更新提示修复、含括号 / 逗号文件名的打开修复、运行环境安装后立即生效（无需重启）

升级说明

  • v1.1.3 用户注意：本版本含 Rust 引擎变更，无法通过应用内热更升级，需下载便携版覆盖安装
  • 便携版：免安装、免管理员，双击即用（建议先退出旧版再覆盖）

校验和（SHA256）

  • Little Claude v1.1.4.exe — f21a1268b8b917aefdb802180067608c71f7ea80c2854464f8a2ca4d96fc6475

v1.1.4 (rebuilt 2026-08-12)

  • fix: context meter (Ctx bar) showing 0% after reopening a session (token stats recovered from session history)

v1.1.4 (rebuilt 2026-08-13)

  • fix: Ctx bar double-counting on DeepSeek-style endpoints — usage where input_tokens already includes the cached share is no longer summed again (semantics-aware formula, auto-detects Anthropic vs DeepSeek semantics, verified against 96/96 usage-log records)
  • fix: auto-compact trigger aligned with the CLI official SDK math (input + cache creation + cache read + output vs threshold)
  • fix: conversation-list status dot — a finished conversation now keeps a steady green dot (running = pulsing green, finished = steady green, error = red), matching the sidebar current-session card

v1.1.4 (rebuilt 2026-08-13, 2nd)

  • feat: automatic 1M context-window detection — no more hard-coded model list. A LiteLLM model table is cached locally (weekly refresh) and queried at session start; any window size (262K / 512K / 1M) is declared to the CLI and drives the Ctx bar / auto-compact thresholds
  • feat: runtime learning — when a session successfully exceeds 900K tokens, that model is recorded as 1M-capable and the declaration takes effect from the next session
  • fix: third-party models no longer fall back to a 200K window; Ctx bar and auto-compact now use the model's real window
Assets 3
Loading