StudyMate：诊断做完以后，下一步到底能做什么？

StudyMate 已经有学习、测验、计划和搭子入口。把它用于自己的雅思备考时，问题变得具体：完成诊断、看过反馈以后，我现在到底能开始哪项训练？

这篇记录 2026 年 10 月的本地迭代，沿用既有项目继续改进。公开仓库是较早快照，不能用它代表这些本地新增能力。我确定使用场景和验收要求，AI 编程工具完成实现与检查。

分配了分钟数，还缺少练习本身

原计划主要保存分科分钟和每周重点。它能给出安排，却没有为每个时间块配齐材料、产出和反馈路径。当天的可用时间又可能临时变化，要求先完成全部诊断或填写稳定日预算，会把开始练习推迟。

我们先把入口收窄到本次时间和环境。查看建议、准备与取消不开始计时，明确开始才创建作答。四科诊断仍可推进，但日常行动不必等待四科齐全。

随后只接第一种写作段落修改训练：接受任务、写修改稿、提交、看反馈、复盘，再看下一项。这个范围能够逐步验收，不必先重做整个学习系统。

模型失败后，不能要求我再写一次原稿

真实写作评估的有界测试出现过截断与超时；另一次服务端返回成功，也不足以保证反馈完整可用。

因此，首次提交后先冻结原稿，评估状态单独保存。刷新不能丢草稿或把原稿改成修改稿。模型失败时显示待评估，允许显式重试，也保留人工反馈入口。合成 Provider 检查失败、恢复、缓存和并发行为；这些检查说明软件怎样处理故障，不说明供应商评分质量。

浏览器还发现提交按钮被搭子气泡遮住。这个问题直接挡住训练，修复时调整遮挡关系，再正常点击验证，没有用强制点击绕过。

同题改好了，不代表诊断成绩提高了

修改稿已经看过反馈，不能回写成首次诊断成绩。我们保存两者关系，同时保留首次作答，让后续复盘知道改进发生在什么帮助之下。

当前只接通一种写作训练。四科持续训练材料、陌生题复测、考官校准与长期使用还未完成。下一阶段应先检查反馈能否完整返回，再观察一次真实训练是否让下一次行动更清楚。

本轮依据本地产品框架、验收记录和实现核对，未调用模型，也未重跑完整产品。历史真实请求的失败继续保留，软件测试结果不作为提分证据。公开仓库仍用于说明项目来源与早期结构。

StudyMate: turning diagnostic feedback into a task someone can start

StudyMate already had study, quiz, planning, and companion interfaces. Using it for my own IELTS preparation exposed a practical gap: after diagnosis and feedback, which exercise could I start now?

This article covers the October local iteration. The public repository is an older snapshot and does not establish these additions. I set the use case and acceptance requirements; AI coding tools implemented and checked them.

Minutes need an executable exercise

The plan mainly stored subject allocations and weekly priorities. Those blocks still lacked a complete route through materials, work, and feedback. Requiring all four diagnostics or a stable daily budget also delayed starting when today’s availability could vary.

We narrowed the entry to the current time and environment. Preparation and cancellation do not start a timed attempt; explicit acceptance does. We then connected one writing-paragraph revision workflow rather than rebuilding the whole study system first.

An assessment failure must preserve the original

Bounded live assessment attempts encountered truncation and timeouts. A successful HTTP response alone did not establish usable feedback.

The first submission therefore freezes the original, while assessment has a separate saved state. Refreshes preserve drafts. Failures remain pending, with explicit retry and manual feedback available. Synthetic-provider checks cover recovery, caching, and concurrency; they do not measure the provider’s scoring quality.

Browser testing also found that the companion bubble covered a submit button. We fixed the obstruction and verified a normal click instead of forcing the test through it.

Assisted revision is a separate observation

A revised paragraph has benefited from feedback. It cannot overwrite the diagnostic score as though it were an unaided first attempt. We retain the original and its relationship to the revision.

Only one writing-training type is connected. Four-skill materials, unfamiliar-task retesting, examiner calibration, and sustained use remain incomplete. Complete feedback is the next engineering requirement; a real training session is the next product check.

For this article I reviewed local plans, acceptance records, and source without calling models or rerunning the product. Historical live failures remain part of the account. Tests do not establish learning gains. The public repository documents the project’s origin and earlier structure.