You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Questions now cover the whole material, and you can see it. A long material (a transcript of several recordings, a book) used to get about as many knowledge points as questions you asked for, so most of it was never asked about, and the only trace was a dashed circle. Now every material is split into sections (a recording's parts, PDF pages or chapters, headings), and each section is shown as covered, planned but not produced (with the real reason) or never planned: on the 资料 row, in the reader's outline (with a filter for what is not covered), on the draft page (with a bar per recording and the list of uncovered sections, each opening the reader at it), in the Jobs console and in the 做这几页的题 panel. The reader outline lists recordings and their parts only, no longer a 英文原句 / 中文对照 entry under every part.
Coverage strength instead of a question count. 创建题组 asks for 精简 / 标准 / 完整 and says before running what it means (about how many questions, how many sections, how many rounds, the estimated tokens and calls). Longer and more important sections get more questions: one cheap call per ~20 sections rates their importance with a reason you can read on the draft page. The server enforces it: a planning call sees only its assigned sections and each section's quota, and what comes back is checked against them. The 30-question limit is gone for the total (a custom number up to 500; each round is still at most 30 questions); only the original language of a bilingual transcript counts as material.
Long runs you can pause. A run executes its rounds one after another on the same draft; pause between rounds, continue, stop and keep what you have, or let it run to the end (自动补到完整, on by default for 标准 and 完整). A run interrupted by a restart comes back as "Interrupted" with Continue from the next round. The time limit applies to each round, not to the whole run, and you can set a spending cap. The Jobs console shows 「第 3/12 轮 · 覆盖 31% · 已用 … · 预计还要 …」 and the rounds.
One action for what is missing. 「为没覆盖的部分补题」 replaces 继续补齐 N 题 and 补 N 题 (draft page, home, Jobs console, also for archived tasks): it targets exactly the uncovered sections, reuses what failed parts had already planned, keeps approved questions and never changes the requested count; its estimate is shown before it starts.
No batch is dropped silently. A malformed or cut-off review reply is asked again (a cut-off one only for the checks that are missing) instead of dropping the whole batch; a part that still failed keeps its written questions and is retried in the fill rounds; every "why no questions" reads the same everywhere and is the real reason, not a raw English error.
Questions count once per document. The 资料 row and the reader count the same questions: all versions of a document, the same recording imported separately, and the single recordings of a merged one ("含旧版本 / 同一录音的其他资料 / 分录音资料"). A material whose questions are only in drafts says 「还没发布 · 草稿里有 N 题」. Fixed a bug that could count another material's questions for a recording.
Jobs console. Selecting finished tasks, archiving and deleting in bulk work as designed (the tooltip is whole, the selected rows are readable in every accent, Esc and keyboard work); a task interrupted by a restart says 已中断 instead of 失败; the header no longer squeezes the task name.
Install from GitHub works. The DSH plugin market installs from the GitHub repository, which has no built interface; it is now built on install. The repository is described and tagged dsh-plugin so it can be found there. The recommended install is still the release package.
The numbers and states now tell the truth. One shortfall (questions kept, sections still without questions, what the next round does) is shown the same way on the home banner, the draft row, the Jobs console and the draft page, and progress is only 100% when nothing is missing. A refused or invalid key is one plain sentence with 去配置模型 and 接着做, not the provider's English. One correct answer reads 答对了 (学习中), not 已掌握. 小节 always means a coverage unit and 批次 a generation batch. Each state has one main action; a section that keeps failing is no longer retried every round and is listed with its reason; a last round that covers nothing is followed by a fill round before the run stops. 精简 says it stops at about 12% on purpose.
Archify, recommended and supported. The 知识骨架 page recommends the open-source Archify plugin (MIT, by tt-a1i, not affiliated) for drawing interactive HTML diagrams, with the DSH install command; one button hands the outline to the conversation, and the HTML file the agent saves is attached to the outline and opened in an isolated window (scripts run, nothing can reach StudyHub). See docs/companions.md.
Update check every 6 hours while the plugin changes quickly (it was 12).