Changelog
Every release, with what changed and what to download.
v0.1.6
v0.1.6 — 产物完整性校验与供应商请求头 / Artifact verification and provider request headers
2026-10-14 · v0.1.5...v0.1.6
中文
新增
- 产物完整性程序化校验:回合结束后不再依赖模型自查,而是解析本轮工具调用,对照数据库记录与磁盘上的实际字节,核对图表、数学图形和文件是否真正落盘。工具调用失败、调用成功但产物实体缺失等情况都会被发现并自动重试/补生成;产物失败卡片与图表面板的「文件缺失」行均提供手动「补生成」入口,可按正文标记或工具调用定位。
- 每个模型供应商可配置自定义请求头:控制台的供应商编辑页可从 + 列表逐个添加两项 —— 自定义 User-Agent(留空保存为默认
ilearnassist/1.0),以及由运维指定名称的会话标识头(名称无默认、添加后必填,不能使用 authorization 等保留名)。会话 id 在会话创建时生成、对旧数据惰性补齐,并随该会话的所有请求(主循环与全部带外调用)发送,便于支持 OpenCode Go 这类按会话路由/缓存的兼容网关。 - 快捷回复新增「有问题,检查一下」。
修复
- 工具调用返回
Tool error时仍会占用正文里对应的产物标记,导致自动补生成永远不触发 —— 失败的调用不再匹配标记,补生成按预期发生;被后续调用取代的旧调用不再渲染卡片。
升级:无破坏性变更。providers 增加 user_agent/session_header、sessions 增加 provider_session_id 列,均在启动时自动补齐;schema 版本不变(5),无需改配置。
English
Added
- Programmatic artifact verification: after a turn, the server no longer relies on the model to self-report — it parses the turn's tool calls and checks the database rows and the actual bytes on disk to confirm that diagrams, math figures and files really landed. Failed artifact calls, and calls that succeeded whose entity is missing, are now detected and retried/re-generated automatically. Failed tool cards and the diagram panel's "file missing" rows both offer a manual regenerate control, addressable by the inline marker or the tool call.
- Optional custom request headers per model provider: the provider editor in the admin console adds fields one at a time from a + list — a custom User-Agent (a blank value saves as the default
ilearnassist/1.0), and a session header whose name the operator chooses (no default, required once added; reserved names such asauthorizationare refused). A per-conversation id is minted when the session is created (lazily filled for older rows) and sent on every request that conversation makes — the main loop and all out-of-band passes — which helps with OpenAI-compatible gateways such as OpenCode Go that route or cache per session. - New quick reply: "有问题,检查一下" (something's wrong, check it).
Fixed
- A tool call that returned
Tool errorstill consumed the corresponding inline artifact marker, so auto-repair never fired. Failed calls no longer match markers, repair runs as intended, and calls superseded by a later one render no card.
Upgrade: no breaking changes. providers gains user_agent/session_header and sessions gains provider_session_id, all added automatically on startup; the schema version is unchanged (5), and no configuration changes are needed.
Verified: pnpm typecheck; pnpm test (3372); Playwright (394).
Downloads
- macOS · Apple silicon (M-series)
ilearnassist-0.1.6-mac-arm64.dmg160 MB - macOS · Intel
ilearnassist-0.1.6-mac-x64.dmg167 MB - Windows · x64
ilearnassist-0.1.6-win-x64.exe138 MB - Linux · x86_64
ilearnassist-0.1.6-linux-x86_64.AppImage161 MB - Linux · amd64 (Debian/Ubuntu)
ilearnassist-0.1.6-linux-amd64.deb126 MB
Earlier releases
v0.1.5
v0.1.5 — 智能上下文、用户偏好与章节位置 / Smart context, user preferences and chapter positions
2026-10-02 · v0.1.4...v0.1.5
中文
新增
- 实验性智能上下文:composer 里一键开启后,每轮只携带最新 2 条历史消息加一段指引,而非全量历史或压缩总结;即时生效,关闭即恢复原上下文,消息始终全部保留。
- 用户偏好记忆:Agent 可通过
ila_save_preference记录长期要求(如「先给结论」「不要用表格」),也可以在消息里划选文字后用「作为用户偏好」一键提取。持有偏好的会话自动注入系统提示,无需开关;新规则可取代旧规则,用户偏好面板可查看与删除。 - 产物卡片内嵌到正文:图表、数学图形和文件卡片可由模型通过正文标记精确插入到对应段落;未写标记时按内容偏移自动回退位置。复制、检索、标题生成等纯文本场景会自动去除标记。
- 缺失产物自动补生成:若正文中的产物标记没有对应的工具调用(模型写了标记却没真正画图/写文件),回合结束前会自动发起一次补生成调用,只绑定标记所指的工具并沿用标记的名称,行、文件、面板与脉络归属全部走正常路径;已有文件不会被自动覆盖,停止/挂起的回合跳过,补生成失败则标记保留为悬空。悬空的产物位上提供「补生成」按钮可手动重试(覆盖已有文件需确认)。
- 网页一键收藏:
web_fetch卡片上新增收藏按钮;同时强化指引 —— 回答实际引用了的网页必须在回合结束前收藏进参考资料,否则之后无法重新打开,也无法被后续回合读取。 - 计划章节的开始位置:每个计划节点记录首次开讲与历次返回的位置,节点下出现可折叠的「位置 N」子列表,点击即跳转;章节跳转会携带上下文指引,要求模型用
ila_recall读回相关片段后续讲,而非从头重来。ila_recall新增 around / range 模式,可按消息位置读取一段原文。 - 节点跳过状态细化:从未开始的节点不显示图标,学过一半被跳过的节点显示暂停图标;复制会话(fork)时位置与计划消息会同步处理。
- 会话对话框恢复工具清单:新建会话与会话设置对话框中可直接勾选可用工具(从所选助理初始化、可覆盖),不再只能通过助理间接控制。
- 左侧栏宽度可调:拖动会话列表右缘即可调整(200–480px,键盘方向键可微调),宽度按窗口记忆。
修复
- 创建会话时显式选择「不使用助理」会被忽略,新会话错误继承当前会话的助理(含系统提示)——现已按选择真正不使用助理。
- 无记录调用的表格仍绘制通用标题栏;计划容器在子话题进行中/完成时正确汇总进度。
升级:无破坏性变更。新增 user_preferences、plan_node_starts 两张表及若干列,均随启动自动创建;schema 版本不变(5),无需改配置。
验证:pnpm typecheck 全通过;pnpm test 3298 项;Playwright 393 项。
English
New
- Experimental smart context: a composer switch makes every turn carry only the newest two history messages plus a short guidance block, instead of the full history or a compaction summary. It takes effect immediately; switching it off restores the prior context, and every message is always kept.
- Remembered user preferences: the agent can record standing requirements via
ila_save_preference("lead with the conclusion", "avoid tables"), or you can select text in a message and use "save as preference" to extract one. Conversations that hold a preference inject it into the system prompt automatically — no toggle; newer rules can supersede older ones, and the Preferences panel lists and deletes them. - Artifact cards inline in the prose: diagram, plot and file cards can be placed at the exact paragraph via in-text markers; without markers, placement falls back to recorded content offsets. Markers are stripped automatically from copy, search, titling and other plain-text uses.
- Auto-repair of missing artifacts: when an in-prose artifact marker has no matching tool call (the model wrote the marker without actually drawing/writing), an out-of-band repair call is made before the turn ends — binding only the marker's named tool and pinning its own name, so the row, file, widget and thread take the ordinary paths. An existing file is never overwritten automatically, stopped/suspended turns are skipped, and a failed repair simply leaves the marker dangling. A dangling slot offers a "regenerate" button for a manual retry (overwriting an existing file asks for confirmation).
- Keep a fetched page in one click: the
web_fetchcard now has a keep control, and the guidance is stronger — any page the answer actually drew from must be kept into the conversation's references before the turn ends, or it cannot be reopened or read by later turns without fetching it again. - Plan chapter start positions: each plan node records where it was first taught and every later return, rendering a collapsible "position N" sub-list that jumps to each spot. Chapter jumps carry context guidance telling the model to read the earlier stretch via
ila_recalland continue rather than restart.ila_recallgains around / range modes for reading a message position or stretch. - Richer skipped-node states: never-started nodes show no icon, partially-studied skips show a pause icon; forks carry positions and remapped plan messages along.
- Tool checklist back in session dialogs: the new-session and session-settings dialogs let you tick the available tools directly (seeded from the chosen assistant, overridable), instead of controlling them only through an assistant.
- Resizable left sidebar: drag the conversation list's right edge (200–480px, arrow keys nudge it); the width is remembered per window.
Fixed
- An explicit "no assistant" choice when creating a session was discarded, and the new session wrongly inherited the current conversation's assistant, system prompt included. It now genuinely uses none.
- A table with no recorded call still draws the generic title bar; plan containers roll progress up from their children correctly while underway and when completed.
Upgrade: no breaking changes. Two new tables (user_preferences, plan_node_starts) and new columns are created automatically on startup; the schema version is unchanged (5), and no configuration changes are needed.
Verified: pnpm typecheck; pnpm test (3298); Playwright (393).
v0.1.4
v0.1.4 — 上下文压缩与历史召回 / Context compaction and history recall
2026-09-29 · v0.1.3...v0.1.4
中文
新增
- 手动上下文压缩:composer 一键把此前历史压缩为一段总结,后续只发送「总结 + 压缩位点之后的原文」,显著节省输入 Token。
- 随时恢复全量上下文:消息全部保留,恢复只是把指针移回;再次压缩用「旧总结 + 新消息」滚动生成,成本与新增内容成正比。
- 预览弹窗:查看当前模式(全量/压缩)、压缩位点、覆盖条数与总结正文。
- 安全网:压缩期间暂停发送(草稿保留);失败提示明确、不阻断、可直接重试;有等待回答的卡片时禁止压缩。
ila_recall历史查询工具:Agent 可主动检索本会话历史 —— 最近 N 条 / 关键词搜索,均支持分页;仅当前会话、仅消息正文;工具描述内置使用场景提示。- 用量统计:上下文占用 popover 新增「本会话累计」(调用数/输入/输出/合计)。
升级:无破坏性变更,新表与新列随启动自动创建,无需改配置。
验证:pnpm typecheck 全通过;pnpm test 3131 项;Playwright 375 项。
English
New
- Manual context compaction: one click in the composer compresses the earlier history into a summary; later turns send only the summary plus the messages after the compaction point, cutting input tokens sharply.
- Restore the full context at any time: every message is kept — restoring just moves the pointer back. A re-compaction folds the previous summary with the new messages, so its cost stays proportional.
- Preview dialog: the current mode (full/compacted), the point, how many messages are covered, and the summary itself.
- Safety net: sending is paused while a compaction runs (the draft is kept); a failure is reported clearly, blocks nothing, and is retryable; compaction is refused while a card is waiting for an answer.
ila_recalltool: the agent can look up this conversation's history on its own — recent N messages or a keyword search, both paged; current conversation and message text only; the description teaches when to reach for it.- Usage: the context popover now shows this conversation's cumulative spend (calls/input/output/total).
Upgrade: no breaking changes; new tables and columns are created automatically on startup, and no configuration changes are needed.
Verified: pnpm typecheck; pnpm test (3131); Playwright (375).
v0.1.3
✨ 新功能
从任意消息创建会话分支
- 每条消息的悬浮操作新增「分支」:以该消息为节点,把当前会话完整复制为一个新会话
- 复制的不只是消息:对话参数(模型、系统提示词、工具白名单)、计划及全部历史版本、测验题、脉络、笔记、图表与表格、洞察、资料引用、
@引用、组件安装、题号计数,以及会话自己的文件目录 - 新会话默认名为「原名 · 分支」,可在对话框中修改
- 原会话有回复正在生成时会拒绝分支(避免复制半截回复);分支只读取原会话,不受会话写入锁限制
测验题可以删除
- 已出的题目可删除:从测验面板和智能体查询中移除,且不会再被要求补答
- 已作答的记录不会销毁:答案、判定、反馈和时间都保留,题卡置灰显示「已删除」,而不是继续提供已被拒绝的补答
- 重复删除无副作用;删光一张正在作答卡片上的题目后,卡片仍可正常关闭
- 题目详情窗口新增「定位」,可跳回并关闭对应题卡
函数图:智能体现在会画坐标图
- 新工具
ila_plot:支持函数、隐式方程、参数曲线、点、线段、多边形、向量、圆和标签 - 表达式走封闭的共享白名单(写入前与渲染前各校验一次),只传数据、不执行任何代码
- 函数图是「图表」面板的第三种内容(图 / 表 / 函数图),并享受与 Mermaid、表格相同的整套能力:卡片、查看器缩放与下载、追问、笔记、定位、
ila_query、脉络归类、洞察、分支
🔄 改进
会话锁:离开即释放,被占用可强制接管
- 关闭或刷新页面、退出登录、切换工作区都会立即归还编辑权,不再等 2 分钟租约过期;浏览器前进后退恢复(bfcache)会自动重新占用,回到前台会立即重新校准锁状态
- 会话被其他设备占用时,只读提示旁新增「强制占用」按钮(带二次确认);被抢占的设备在下一次写入被拒时立即切换为只读,不再长时间停留在可写状态
🐛 修复
被占用会话中发送失败的消息不再残留
- 在锁定会话里发送被拒后,消息不再留在对话中(此前只有前端可见、刷新后才消失);文字与已选的附件、引用会退回输入框,不会丢失
v0.1.2
No release notes were written for this version.
v0.1.1
No release notes were written for this version.
v0.1.0
首个公开版本。交互式学习助理 是一个运行在你自己电脑上的 AI 学习助手。它不是又一个聊天框: 会制定学习计划、逐项讲解、出题检验,并把笔记、图表、参考资料整理成看得见的面板。所有数据都在 你选择的文件夹里,只有你自己配置的大模型接口会收到你的对话。
下载
| 系统 | 文件 |
|---|---|
| macOS(Apple 芯片,M 系列) | ilearnassist-0.1.0-mac-arm64.dmg |
| macOS(Intel 芯片) | ilearnassist-0.1.0-mac-x64.dmg |
| Windows 64 位 | ilearnassist-0.1.0-win-x64.exe |
| Linux 64 位 | ilearnassist-0.1.0-linux-x86_64.AppImage 或 ilearnassist-0.1.0-linux-amd64.deb |
安装前请先读这一段
安装包没有购买代码签名证书,所以系统会拦一下 —— 这不是文件损坏:
- macOS:在应用上点右键 → 打开 → 再点一次"打开"。只需一次。
- Windows:在 SmartScreen 提示里点 更多信息 → 仍要运行。
- Linux:AppImage 下载后先
chmod +x,再运行。
装好之后
- 启动应用,控制面板出现。
- 选一个数据文件夹。面板会建议
~/ilearnassist(你的主目录下),点一下就用它;也可以 自己挑一个位置。这里存放你的数据库、工作区和上传的文件。 - 创建超级管理员:填一个用户名和密码。这是这台机器上权限最高的账号,请记住它。
- 服务器会自动启动。点 打开应用,用刚才的账号登录。
- 填一个模型 API Key:左侧菜单 → 平台管理 → 模型服务,选一家你已经开通的服务商填入。 填好之前,模型列表是空的 —— 助手需要一个大模型才能工作。
- 开始学习:回到首页,新建一个工作区,进去后新建对话,助理选 引导学习 · Guided Learning,然后告诉它你想学什么。
这一版有什么
- 内置「引导学习」助理,以及七个学习控件:计划 / 测验 / 脉络 / 笔记 / 图表 / 思考 / 参考资料
- 内置 11 家模型服务商配置(DeepSeek、通义千问、智谱 GLM、Kimi、MiniMax、OpenAI、Gemini 等), 填一个 API Key 就能开始
- 资料库与
@引用、文档解析(PDF / Office 本地提取,可选云端解析扫描件)、文件管理 - 数据库迁移框架:以后升级自动迁移,而且在动手之前先给数据库做一份快照
- 数据与程序分离:卸载应用不会删除你的数据
- MIT 许可证,可自托管
English
First public release. Interactive Learning Assistant is a self-hosted AI study assistant that plans a syllabus, teaches one item at a time, quizzes you, and keeps your notes — every byte stays in a folder you choose.
The installers are unsigned, so your OS will warn you: on macOS right-click the app → Open →
Open again; on Windows click More info → Run anyway; on Linux chmod +x the AppImage.
Then: pick a data folder → create the superadmin → add a model API key (Platform admin → Model services; the model list stays empty until you do) → create a workspace, start a conversation, and choose the built-in Guided Learning assistant.