dsh-agy-link 0.4.67 → 0.4.70

package/CHANGELOG.md

@@ -1,5 +1,129 @@
 # Changelog
 
+## 0.4.70 (2026-10-25)
+
+### English
+
+- **Fix: the 5-hour quota row showed the weekly numbers (2026-10-22 report).**
+  - `fetchAvailableModels` returns a per-model `quotaInfo` with NO window label. For some accounts it carries the 5-hour bucket, for others the WEEKLY bucket: live `daoxzeex@gmail.com` reported `0.1736 / reset 2026-10-24T18:19:18Z` on every Gemini model, which is exactly the weekly bucket of `retrieveUserQuotaSummary`.
+  - The merge took "the lowest remaining across all sources" and filed it into the family 5-hour slot, so the GUI painted `5-hour quota 19% / Weekly 19%` with the SAME weekly reset on both rows (the filed screenshot) while the account actually had 82% of its 5-hour quota left.
+  - The reset instant now decides the window (`sameQuotaWindow`, 5-minute tolerance): a reading that matches the summary weekly reset tightens ONLY the weekly row; one that matches the 5-hour reset tightens the 5-hour row; an unmatchable reading keeps the old conservative behaviour. Eligibility is unchanged — the weekly row still drives the 5% drain check.
+- **Fix: a queued turn no longer looks frozen for 15 minutes.**
+  - Windows has ONE machine-global agy credential slot, so sessions serialize: a healthy long run in one workspace made two other sessions sit with a completely empty transcript for `dispatchTimeoutMinutes` (15m) and then silently re-queue for another cap — indistinguishable from a hang (2026-10-22: the promo-tcn sessions, while a 15-minute run in another project held the slot).
+  - The wait now streams a session-visible `[agy] waiting for the agy credential slot …` line into the transcript, with an escalating cadence (20s → 60s → 3m → 9m) so even a 15-minute queue leaves a handful of lines. Re-queues after a cap are announced too.
+  - A generator abandoned while waiting (Stop / stream closed) now hands a late grant back instead of leaking a concurrency slot.
+- **Fix: the machine-wide credential slot is held only until agy has READ it — different accounts now run concurrently.**
+  - Measured on agy 1.4.3 with a live 14-minute run: print mode resolves its credential ONCE at startup (`printmode.go:403 "not authenticated, trying silent auth"` → `:405 "silent auth succeeded"` at +3.2s) and performs zero further token/auth work for the rest of the run. Holding the Windows slot for the whole run was pure serialization, and it is what parked the 2026-10-22 promo-tcn sessions for 15 minutes behind another project's run.
+  - `armSlotHandoff` releases the lease on the child's FIRST output line (45s ceiling for a slow startup, run-end fallback, idempotent), while the identity is still installed under the mutex immediately before each spawn — so a later switcher cannot change what an already-started run read.
+  - `JETSKI_OAUTH_TOKEN` (the only env-based credential route in the binary) was probed with a bogus token: the run still authenticated from the keyring, so true per-process injection remains unavailable upstream. Recorded in KNOWN-GAPS with the measurement.
+- **Fix: a rejected credential-slot write now refuses the run instead of billing the account that happens to be installed.**
+  - The Windows write is a PowerShell cold start that compiles a C# P/Invoke shim; its old 8-second ceiling made it time out and return `false` **without touching the slot** on a loaded machine (reproduced 2026-10-23: delete then write reported false and left the slot empty). All four install sites ignored that boolean, so the next spawn authenticated as the PREVIOUS account — silent cross-account quota spend, the exact corruption class the module exists to prevent.
+  - Install sites now verify the write (one retry, then `SlotUnavailableError` → the turn fails visibly); teardown/restore paths log loudly instead of failing a finished turn. Writes get a 30s ceiling; reads stay at 8s because they sit on the poll path.
+  - New test: a write that always reports false refuses the run, leaves the slot untouched and does not leak the mutex.
+
+### 中文
+
+- **修复：5 小时额度行显示成了周额度（2026-10-22 反馈）。**
+  - `fetchAvailableModels` 的 per-model `quotaInfo` 不带窗口标签：有的账号返回 5 小时桶，有的账号返回**周**桶（实测 `daoxzeex@gmail.com` 每个 Gemini 模型都返回 `0.1736 / reset 2026-10-24T18:19:18Z`，正是 `retrieveUserQuotaSummary` 里的周桶）。
+  - 旧合并逻辑「取所有来源里最低的余量」把它塞进了 5 小时行，界面于是显示 `5 小时额度 19% · 周额度 19%`、两行 reset 完全相同（反馈截图），而该账号 5 小时其实还剩 82%。
+  - 现在由 reset 时刻判定窗口归属（`sameQuotaWindow`，5 分钟容差）：命中周 reset 的读数只收紧周行；命中 5 小时 reset 的收紧 5 小时行；无法判定的读数保持原有保守行为。可用性判定不变——周行依旧参与 5% 耗尽检查。
+- **修复：排队中的回合不再看起来卡死 15 分钟。**
+  - Windows 只有一个全局 agy 凭据槽，会话之间必须串行：另一个工作区一次正常的长任务就能让其它会话的转录区**整整 15 分钟一片空白**（`dispatchTimeoutMinutes`），随后再静默重排一轮——和卡死无法区分（2026-10-22 的 promo-tcn 两个会话就是被另一个项目 15 分钟的运行挡住）。
+  - 等待期间现在会向会话流里写入可见的 `[agy] 正在等待 agy 凭据槽 …` 提示，节奏递增（20s → 60s → 3m → 9m），15 分钟排队也只留几行；超时重排同样会写明。
+  - 等待中被 Stop / 关流放弃的请求，现在会把迟到的许可归还，而不是漏掉一个并发槽。
+- **修复：全局凭据槽只占用到 agy 读完凭据为止——不同账号现在可以真正并发。**
+  - 实测 agy 1.4.3 一次 14 分钟的运行：print 模式只在启动瞬间解析一次凭据（`printmode.go:403` 未认证 → `:405` silent auth 成功，+3.2 秒），之后 14 分钟**再无任何 token/auth 操作**。把 Windows 全局槽占满整个 run 纯属多余串行——2026-10-22 的 promo-tcn 两个会话正是被另一个项目那 14 分钟的运行堵了 15 分钟。
+  - 新增 `armSlotHandoff`：在子进程输出**第一行**时立刻交还租约（启动慢则 45 秒上限、run 结束兜底、幂等），而身份安装依旧在每次 spawn 前、在互斥锁内完成，因此后来的切换不可能改变已启动 run 读到的凭据。
+  - 二进制里唯一的进程级凭据入口 `JETSKI_OAUTH_TOKEN` 已用假 token 实测：仍然从系统 keyring 认证成功，说明上游还不支持真正的进程级注入；该结论连同测法已写入 KNOWN-GAPS。
+  - 补充实测：**把全局凭据条目删掉后，agy 确实会改读家目录凭据文件**（假 token → `UNAUTHENTICATED 401`）。但 keyring 在写入侧同样优先（`Failed to save token to keyring, falling back to file`），任何一次成功运行都会把全局条目写回来，所以"永久空槽 + 纯文件隔离"仍不成立；凭据隔离目前只是读取侧能力。
+- **修复：凭据槽写入被拒绝时，拒绝启动，而不是用"当前碰巧装着"的账号去计费。**
+  - Windows 侧写入是一次 PowerShell 冷启动（Add-Type 编译 C# P/Invoke），旧实现的 8 秒上限在机器繁忙时会**超时并返回 false，而槽位根本没被改动**（2026-10-23 实测：删除后再写入返回 false，槽位保持空）。四个安装点都没检查这个布尔值，于是紧接着的 spawn 会以**上一个账号**的身份认证 —— 静默跨账号扣额度，正是本模块要防的那类污染。
+  - 现在安装点会校验写入（失败重试一次，仍失败则抛 `SlotUnavailableError`，让这一轮可见地失败）；收尾/恢复路径改为大声告警而不毁掉已完成的轮次。写入超时提到 30 秒，读取保持 8 秒（在轮询路径上）。
+  - 新增测试：写入恒返回 false 时必须拒绝运行、槽位保持原状、互斥锁不泄漏。
+
+## 0.4.69 (2026-10-23)
+
+### English
+
+- **Feat: Google risk-control hardening and CLI-first quota architecture.**
+  - Implemented CLI-first quota detection via official `agy -p /usage`: 100% genuine Go TLS fingerprint and client handshake, eliminating Node.js-level direct REST calls (`v1internal:retrieveUserQuotaSummary`) that trigger JA3/JA4 fingerprint mismatch warnings.
+  - Purged internal test endpoints (`daily-cloudcode-pa.sandbox.googleapis.com` and `autopush-cloudcode-pa.sandbox.googleapis.com`) from `AGY_ENDPOINTS` to prevent sandbox probe alarms on consumer accounts.
+  - Token refresh is fully delegated to the native `agy` CLI Go runtime; Node.js host process no longer issues direct refresh requests to `oauth2.googleapis.com/token`.
+  - Background quota polling interval widened from 15m to 60m (`quotaPollIntervalMs: 60m`), and account polling incorporates a 2-second stagger sleep to avoid simultaneous process bursts.
+- **Feat: Full concurrency on Windows credential slot via global primary account alignment (Scheme A).**
+  - Resolved Windows single-slot mutex serialization: concurrent sessions align to the current primary healthy candidate, hitting the `parallelSameAccount: true` fast path for 100% lock-free concurrency across sessions.
+  - Single-direction migration to the next candidate only occurs when the active account encounters a 429 quota exhaustion.
+- **Feat: Two-tier dual-loop test framework aligned with academic-agent & test suite speedup.**
+  - Introduced `pnpm run test:fast` (powered by `scripts/fast-test.mjs`) for inner TDD vibe loops (0.5s~5s), running only Git-affected test files or specified module keywords (`pnpm run test:fast <keyword>`).
+  - Tuned simulated asynchronous delay timers in `test/adapter.test.ts` and `fake-agy.mjs` from 5000ms/8000ms down to 300ms/400ms, speeding up adapter unit test execution by ~4x.
+- **Feat: Scroll guard & tool view expansion state persistence.**
+  - Added `scroll-guard.ts` to suspend DOM mutations and heavy scans (decisions, images, reasoning, mermaid) during active scrolling and smooth momentum settling.
+  - Preserved user expanded/collapsed tool card states across TanStack Virtualizer dynamic remount cycles.
+
+### 中文
+
+- **新功能：针对 Google 敏感风控强化防护与 CLI 优先配额体系。**
+  - 配额查询全面转向官方 `agy -p /usage` 解析：100% 走官方原生 Go 二进制握手与正规 TLS 指纹，消除 Node.js 层使用 `undici` 直接裸调 Google 内部 `v1internal` 接口带来的 JA3/JA4 客户端指纹断层风险。
+  - 从 `AGY_ENDPOINTS` 中彻底移除 `daily-cloudcode-pa.sandbox.googleapis.com` 与 `autopush-cloudcode-pa.sandbox.googleapis.com` 内部测试端点，杜绝异常沙箱端点探测对正式 Google 账号的风控隐患。
+  - OAuth 令牌刷新完全交由 `agy` 官方 Go 运行时在日常使用中原生完成，Node.js 宿主停止主动向 `oauth2.googleapis.com/token` 发送换票请求。
+  - 默认后台巡检周期放宽至 60 分钟（`quotaPollIntervalMs: 60m`），多账号后台轮询增加 2 秒错峰休眠，避免瞬时并发拉起 CLI 进程造成的系统负载颠簸。
+- **新功能：Windows 凭证槽位并发性能全面解锁（方案 A：全局主导账号收敛）。**
+  - 解决 Windows 系统下因全局唯一定制凭证槽位 `gemini:antigravity` 导致的跨 Session 串行卡顿问题：所有会话默认对齐到号池主导活跃健康账号，命中 `parallelSameAccount: true` 快速通道，实现跨会话 100% 无锁全并发执行。仅当主导账号 429 配额耗尽时，全局指针单向整体迁移至下一账号。
+- **新功能：对齐 academic-agent 的两级双环极速测试架构与用例调优。**
+  - 新增 `pnpm run test:fast` 极速内环测试工具（`scripts/fast-test.mjs`），实现基于 Git 变动的秒级精准回归（0.5s~5s）以及按模块关键字直跑（`pnpm run test:fast <keyword>`）。
+  - 优化 `adapter.test.ts` 与 `fake-agy.mjs` 中的异步等待延迟（由 5000ms/8000ms 降至 300ms/400ms），单测执行速度提升近 4 倍，保障日常开发极速反馈心流。
+- **新功能：页面平滑滚动守卫与工具卡片展开状态持久化。**
+  - 增加 `scroll-guard.ts` 滚动守卫机制，在用户主动滚屏及平滑滑动期间挂起 DOM 探测（decisions/images/reasoning/mermaid），彻底杜绝滚屏时内容突变抖动与闪烁。
+  - 工具卡片折叠/展开状态跨 TanStack-Virtual 虚拟列表重新挂载自动持久化保留（`expandedToolBlockKeys`）。
+
+## 0.4.68 (2026-10-23)
+
+### English
+
+- **Feat: Dynamic model picker reflects active account eligibility (Gemini 4 Argon isolation).**
+  - Accounts in the pool now discover and cache models independently (`ModelCatalog` keyed by `accountId`).
+  - When switching the active account in the UI or pool, the DSH composer model picker automatically refreshes (`llm/adapters-updated`).
+  - If the active account lacks Argon access (free/standard tier), `gemini-4-argon` is dynamically excluded from the picker to prevent generation errors. When an account with Pro/Ultra subscription is active, Argon models are immediately restored.
+- **Fix: Eliminate page scrolling and violent screen flickering in Antigravity client.**
+  - Root cause: In `src/client/decisions.ts`, decision chip toolbar was repeatedly injected into TanStack-Virtual dynamic measurement containers because container-level marker `data-agy-decisions-scanned` was checked but never set, and re-rendering reconstructed heading nodes. The item height expansion triggered scroll position adjustments, which caused a re-render and re-injection flywheel.
+  - In addition, settings nav row `MutationObserver` watched the entire subtree including `characterData` and re-triggered on its own marker writes without an attribute filter, and account button busy states repainted the entire account list instead of scoping to the target account.
+  - Resolved with container-level idempotence guards (`data-agy-decisions-scanned` written before mutation), pre-checking existing toolbar containers, filtering nav mutation observer strictly to marker attribute changes, and scoping button busy states and shimmer animations to target accounts.
+- **Fix: Send agy CLI User-Agent in quota requests so Argon quota group is returned (PR #54, Fixes #53).**
+  - Google's Cloud Code backend requires a genuine `agy` CLI User-Agent header (format `Antigravity/<version> (darwin; arm64)`) to return the `Gemini 4 Argon` quota bucket in `retrieveUserQuotaSummary`. Without this header, only generic Gemini and 3P groups were returned.
+  - Quota service now detects the installed `agy` CLI version and passes the proper User-Agent header in HTTP requests. When an account transitions or drops dedicated groups, obsolete cooldowns and stale last-known-good entries are cleanly evicted.
+- **Fix: DSH Host 100% CPU busy loop on macOS caused by copying active SQLite WAL files (Fixes #52).**
+  - Root cause: `loadAgyDbData()` previously copied `conversation.db`, `-wal`, and `-shm` to a temporary directory before querying tool step parameters and model thoughts. On macOS (and other BSD-derived libuv platforms), when an active SQLite WAL file is checkpointed or truncated to 0 during `copyFile`, libuv's `uv_fs_copyfile` enters an infinite busy loop in a worker thread, causing the Host process to permanently consume 100% of a CPU core.
+  - Solution: Replaced file copying with direct SQLite reads (`node:sqlite` DatabaseSync with `readOnly: true`, falling back to `sqlite3` CLI with `-cmd ".timeout 2000" -cmd "PRAGMA query_only = ON;"`). SQLite's native WAL architecture provides atomic, consistent MVCC read snapshots concurrently with active writers without blocking or file-copying races.
+  - Added regression test covering active WAL querying, WAL truncation/checkpointing during active use, sqlite3 CLI fallback, and zero-byte database handling.
+- **Fix: Transport failures outlast a real egress outage, and no longer rotate a shared-egress pool.**
+  - Live case (2026-10-21 16:34–16:48): the machine's proxy answered `CONNECT 200` and then killed TLS for ~14 minutes. Every `agy` run on every account died in its startup probe with `Post "https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist": EOF`, which agy wraps as `Eligibility check failed: …`. DSH retried twice inside ~4 seconds and gave up — while the Antigravity desktop's own language server simply kept retrying and recovered.
+  - Transport (`NETWORK`) failures now state their own backoff via `providerRetryAfterMs`: **5s → 15s → 30s** for consecutive failures in one session, reset on success or any non-transport failure, and treated as fresh after 10 minutes of quiet. `providerRetryPolicy` grew to 3 retries with `maxDelayMs: 45s` (the retry plugin *drops* a retry whose requested delay exceeds that cap, so it must sit above the longest transport delay).
+  - A transport failure is only account-scoped when accounts leave through **different** proxies. With a single shared egress (every account on the config/system proxy, including pool accounts seeded with a copy of the system proxy) the retry stays on the current account: the sticky pointer is not advanced, the conversation binding is kept, and the account-rotation budget stays available for real quota/eligibility hops. The in-step failover of a settled recording also skips the budget without changing accounts.
+  - Non-transport retries keep their 2s/4s starting schedule; nothing else about classification (AUTH vs NETWORK vs ELIGIBILITY) changed.
+
+### 中文
+
+- **新功能：模型选择框按当前活动账号资格动态隔离（Gemini 4 Argon 自动感知）。**
+  - 号池内账号支持按账号独立发现并缓存模型目录（`ModelCatalog` 按 `accountId` 隔离）。
+  - 在前端或号池切换当前活动账号后，DSH 输入框模型选择列表即时联动刷新（`llm/adapters-updated`）。
+  - 若当前活动账号无 Argon 权限（标准/免费层），输入框模型列表中自动剔除 `gemini-4-argon`，防止误选调用报错；当切回 Pro/Ultra 订阅账号时，Argon 模型自动呈现。
+- **修复：解决 Antigravity 插件设置及历史页面来回滚动疯狂闪屏问题。**
+  - 根因定位：在 `src/client/decisions.ts` 中，决策按钮增强器容器级标记 `data-agy-decisions-scanned` 此前只读未写，且在 TanStack-Virtual 动态虚拟列表中，重新挂载渲染会重建 heading 节点。工具条反复注入导致项高度变化，进而触发虚拟列表重算 scrollTop 并再次触发重渲染注入飞轮。
+  - 此外，设置导航图标的 `MutationObserver` 监听范围过广且在修改属性时自身被反复触发，且操作账号按钮的全局 busy 状态导致 40+ 个元素属性全量刷新并引发重排重绘。
+  - 修复方案：在变更 DOM 前对容器写入 `data-agy-decisions-scanned` 并严格预检容器已有工具条；导航 Observer 限制仅监听对应属性变更（`attributeFilter`）并剥离 `characterData`；按钮 loading 状态与额度进度条 shimmer 动画改为账号级别作用域隔离。
+- **修复：配额请求中上报 agy CLI 原生 User-Agent 以获取 Argon 额度分组（合并 PR #54，修复 #53）。**
+  - Google Cloud Code 后端要求在 `retrieveUserQuotaSummary` 请求中携带真实的 `agy` CLI User-Agent（格式如 `Antigravity/<version> (darwin; arm64)`）才会返回 `Gemini 4 Argon` 独立额度桶。缺少该头时后端仅下发普通 Gemini 与 3P 组。
+  - 配额探测服务现在自动探测本地 `agy` 版本并携带该头发送请求；当账号升级或移出专用桶时，旧版冷却记录与失效缓存自动淘汰。
+- **修复：macOS 下复制活跃 SQLite WAL 时触发 libuv `copyFile` 忙循环导致 DSH Host 持续 100% CPU（修复 #52）。**
+  - 原因分析：`loadAgyDbData()` 此前在读取工具参数和思考内容时，顺序将 `conversation.db`、`-wal` 和 `-shm` 复制到临时目录。在 macOS（以及其他基于 BSD 的 libuv 平台）上，当活跃 WAL 文件在复制过程中被 SQLite checkpoint 截短为 0 字节时，libuv 的 `uv_fs_copyfile` 工作线程会陷入无限忙循环，导致 DSH Host 进程持续耗尽单核 100% CPU。
+  - 修复方案：完全移除临时目录与文件复制逻辑，改为直接通过 SQLite 引擎一致性读取（优先使用 `node:sqlite` 的 `DatabaseSync` 只读模式并开启 `PRAGMA query_only = ON;` 与 `busy_timeout`，后备使用 `sqlite3` CLI）。SQLite WAL 模式天然支持多读一写并发 MVCC 快照，既消除了文件截短忙循环与临时文件 I/O 争用，又保证了读取一致性。
+  - 补充回归测试：覆盖活跃 WAL 读取、WAL 截短/检查点处理、sqlite3 CLI 后备路径及 0 字节文件防御。
+- **修复：网络抖动时重试不再 4 秒内放弃，共用出口的号池不再被无意义轮换。**
+  - 实测案例（2026-10-21 16:34–16:48）：本机代理对 `CONNECT` 返回 200 后直接掐断 TLS 长达约 14 分钟。每个账号上每次 `agy` 运行都在启动探测阶段死于 `Post "https://daily-cloudcode-pa.googleapis.com/v1internal:loadCodeAssist": EOF`，被 agy 包成 `Eligibility check failed: …`。DSH 约 4 秒内重试两次即宣告失败，而 Antigravity 桌面端自己的 language server 只是一直重试，网络恢复后照常工作。
+  - 传输类（`NETWORK`）失败现在通过 `providerRetryAfterMs` 自带退避：同一会话连续失败按 **5s → 15s → 30s** 递增；成功或出现非传输类失败即归零，静默超过 10 分钟后重新从 5s 起算。`providerRetryPolicy` 提升为 3 次重试、`maxDelayMs: 45s`（重试插件对**超过上限**的请求延迟会直接放弃该次重试，因此上限必须高于最长退避）。
+  - 只有当各账号走**不同代理**时，传输失败才算账号级问题。共用同一出口时（所有账号都走配置/系统代理，含新建号池账号被写入系统代理副本的情况）重试留在当前账号：不推进粘性指针、保留会话绑定、也不消耗本该留给配额/资格跳转的轮换预算；已结算记录的步内失败切换同样不再计数、不再换号。
+  - 非传输类重试仍从 2s/4s 起步；AUTH / NETWORK / ELIGIBILITY 的判定逻辑未做任何改动。
+
 ## 0.4.67 (2026-10-20)
 
 ### English