v1.9.0

v1.9.0 — 账号安全：推送前自检 + 抗风控策略

GitHub 的滥用判定看行为形态，不看意图 —— 哪怕只是在原仓库做个小修复也可能触发。本版把四类高危形态做成了自动检查：

检查	之前	现在
secret 扫描	推上去才被 GitHub 拦	推送前本地扫 OAuth secret / token / 私钥，命中直接拒绝
大文件	50MB+ 靠 GitHub 告警	推送前主动告警，建议 LFS 或剔除历史
失败重试	任何失败都重试 6 次	只对传输错误重试，指数退避（2/4/8/16/30s）；服务端拒绝立即终止
tag 重指	静默 force-push 已发布 tag	需显式 -AllowForceTag

为什么「重试」最危险：同一笔被拒的 push 反复重发，是最典型的滥用信号。很多人（包括我朋友的账号）就是卡在这一步——被拒后不断重试，最后账号受限。

skill.md 新增「账号安全」小节，故障速查表也补了哪些错误不要重试。

v1.8.2

插件卡片显示与版本兼容

• exports["."] 增加 "./package.json"，让插件市场与插件管理器能读到 title/description（原先回落成包名 + 文件路径）
• 新增 engines.dsh: ">=0.1.0"，显式声明兼容下限

无功能变更。bundle.gate / safety.gate / features.gate 全绿。

v1.8.1 - follow account rename

chore: follow the account rename to JiangHuai24

Repository URLs, the git remote and the scrub pattern now point at the new
account. The account store is redirectable so tests stop writing the live
file, and the feature gate uses a scratch path.

v1.8.0 - composer push prompt and accounts

feat: in-conversation push prompt, multi-account, and pre-push scrub

A composer button fills a push prompt into the current input via
inputActions.setDraft (never auto-sends). Accounts are managed by name only:
no tokens are stored, authentication stays in Git Credential Manager and
switching just changes the default identity. A scrub route scans the files a
push would publish for the personal account, tokens, keys and emails; the
repository itself has been cleaned so the published content carries the
project identity rather than a personal handle.

v1.7.0 - safe config writes

feat: apply the same safe patch persistence

v1.6.1 - graceful degradation

fix: never abort activation when skills is unavailable

v1.5.3 - reliable push verification

fix: retry ls-remote and require a real sha before verifying

v1.5.1 - safe children spreading and bundle gate

fix: spread React children safely and adopt the bundle gate

v1.5.0 - setup wizard

feat: settings form plus a step-by-step setup wizard

Adds an editable config form and a three-step wizard modelled on gh auth login
and create-vite: one decision per screen, a progress rail, per-step gating and
a read-only confirmation screen before anything is written.

v1.4.1 - optical icon sizing

style: redraw the arrow to match the other glyph extents