这里记录我们正在做什么。

Update

TAO 发布链路和 workspace 调度稳定性更新

Summary

TAO 这轮把桌面端标准发版、运行目录治理、workspace 容量调度和 center 部署重新固化，重点解决反复打包、磁盘膨胀和容量不足时卡在原宿主机重试的问题。

Progress

1. 固化桌面端标准发版流程：标准构建脚本和发布后校验脚本已经固定下来，减少反复打包时的偏差。
2. 清理运行目录膨胀问题：运行目录不再保留重复 staging 文件，并增加默认运行容量上限，单 workspace 的验证容量也加了明确边界。
3. 修复重复弹窗和 report bridge 问题：当运行根目录不存在时，桌面 runtime control 不再反复拉起 report bridge。
4. 增加容量感知调度：宿主机容量不足时会上报明确失败码，center 会清空原节点绑定，将 workspace 重新置为 pending，并排除失败节点重新调度。
5. 完成 center 侧部署与迁移：服务端已经更新到最新版本，并补齐 workspace_instance 的 failure_code、failure_message、failure_node_id 相关迁移字段。
6. 验证发布结果：本地发布校验通过，center API 能看到容量不足后的 pending workspace 不再继续绑死原节点。

Problems

主要问题是运行目录曾经涨到异常体积，workspace 容量不足时还会在同一台机器上重复尝试，导致用户看到创建流程卡住。另一个问题是桌面端后台辅助进程曾经可能反复弹出窗口，影响稳定使用。

Solutions

这轮把标准构建脚本、发布校验脚本和容量失败码沉淀为固定流程。桌面端负责准确上报容量失败，server 负责记录失败节点并重新调度；如果当前没有可托管节点，workspace 保持 pending，而不是继续消耗同一台机器。

Next actions

下一步继续跑两台机器、两套用户的完整闭环：注册登录、自动创建 Ubuntu workspace、用户之间消息互通、用户到自己 workspace agent 的消息、workspace go agent 到各机器 rust agent 的控制链路，并补齐跨机器容量调度和恢复测试。

Update

XXB desktop question-bank import reached a usable local test state

Summary

XXB now has a running Tauri desktop client and a clearer question-bank import flow that turns uploaded papers into structured questions with stem clues, key points, and per-option explanations.

Progress

1. Started the Windows Tauri desktop client locally and verified that the real window opens.
2. Kept the desktop client pointed at the production entry: https://xxb.life/exam-coach.
3. Added a question enrichment step during bank import so each structured question can carry stem clues, key points, and per-option explanations.
4. Connected both PostgreSQL-backed storage and file-backed local storage to the same enrichment helper.
5. Updated the Exam Coach question-bank page so imported drafts and saved banks use the same formal question card.
6. Changed the import preview to show the backend-enriched questions returned by the import endpoint, rather than only the raw local recognizer result.
7. Added focused checks for the enrichment helper and verified locale JSON validity.
8. Captured a desktop screenshot showing the XXB login page in the Tauri window.

Problems

The frontend-wide lint/typecheck path is still slow in this workspace and surfaces unrelated existing ChatPanel/path-alias issues when run outside the normal Next project context. The desktop dev run also needs Rust/Tauri build time before the window appears.

Solutions

The implementation was kept narrow: enrichment is isolated in one service helper, existing import endpoints continue using the same table shape, and the UI reads optional fields so older imported questions can still render. Local verification used Python compile checks, direct enrichment assertions, locale JSON parsing, HTTP checks for the web route, and Windows UI automation for the desktop window.

Next actions

1. Run a real paper import through the desktop window after login and inspect whether the generated clues and option explanations are useful enough for teachers.
2. Replace heuristic option explanations with model-assisted explanations when a reliable answer and parsing context are available.
3. Add a richer screenshot or short demo video once the logged-in import flow is validated end to end.
4. Clean up unrelated frontend typecheck issues so full CI can verify this page without manual filtering.

Update

Recent Opsera fixes, listed clearly

Summary

A clearer itemized summary of the recent Opsera stability and release work.

Progress

1. SSO status is now more accurate. `sso status` no longer reports the agent as offline just because a health check was slow.

2. The agent now syncs with Xshell automatically. If Xshell is open and a reusable SSH URL exists, status checks can start the agent and attach without needing a manual restart.

3. Stale SSH tokens are handled better. When the Xshell URL changes, Opsera clears its failed-token cache instead of staying stuck on old connection state.

4. DASUSM browser keepalive was fixed. The browser refresh job no longer fails on a PowerShell variable conflict, and minimized browser windows are restored before refresh.

5. Xshell keepalive remains available as a fallback. The helper sends a real Enter on a controlled interval to keep idle terminal sessions active when the environment requires it.

6. Large upload reliability was hardened. `upload-large --sso` now waits for chunk upload, remote merge, final size check, SHA256 verification, and cleanup before treating the upload as successful.

7. Codex skill deployment was updated. The rebuilt `opsera.exe` was copied into the local Codex skill so other agents use the fixed version.

8. Live verification passed. After stopping the agent, `sso status` was able to bring it back, report `opseraConnected:true`, and a live `--sso` command completed successfully.

Problems

The earlier summary was too dense. The technical fixes were correct, but the public update was harder to scan than it needed to be.

Solutions

This update restates the same work as a numbered progress list so readers can quickly see what changed and what was verified.

Next actions

Continue observing long-running SSO sessions and keep using the hardened `upload-large --sso` path for large release artifacts.

Update

SSO stability and transfer reliability update

Summary

Recent work focused on making Opsera's SSO agent more reliable, reducing false offline reports, and hardening large file transfer completion checks.

Progress

We improved the SSO agent so it now synchronizes with live Xshell sessions more reliably. Status checks can start and refresh the agent when a reusable Xshell URL is present, the agent clears failed-token cache when the Xshell URL changes, and health checks no longer confuse a slow status probe with an offline agent. Browser keepalive for the DASUSM page was also fixed and verified. Large uploads were tightened so completion requires chunk upload, remote merge, final size, SHA256 verification, and cleanup.

Problems

The main issue was inconsistent state: Xshell could be open while Opsera still showed offline, or the agent could be healthy while status reported it as not running. Browser keepalive was also being skipped because of a PowerShell variable conflict, and older large-upload behavior could return before the remote file was fully merged and verified.

Solutions

We added automatic SSO status synchronization with Xshell, separated lightweight readiness checks from full health checks, increased the health-check timeout, restored minimized browser windows before refreshing them, and made large uploads block until the final remote artifact is verified. The bundled Codex skill executable was rebuilt and replaced after each fix.

Next actions

Continue observing long-running sessions, keep the Xshell Enter keepalive as a fallback for environments that close idle terminals, and use upload-large for large release artifacts.