dsh-agy-link 0.4.44 → 0.4.49

package/docs/KNOWN-GAPS.md

@@ -2,6 +2,24 @@
 
 Deliberate v1 boundaries and upstream-behavior notes.
 
+## Platform limits
+
+- **Windows: pool credential isolation is implemented by handing the shared
+  slot over (issue #45)** — agy on Windows resolves its Google credential from
+  ONE process-global Credential Manager slot `gemini:antigravity`
+  (go-keyring); `HOME`/`USERPROFILE`/`GEMINI_CLI_HOME` move files only. Rather
+  than pretending the pool is isolated, the plugin now owns the slot: the
+  primary credential is kept in a durable, identity-verified store
+  (`agy-accounts/.primary-credential.json`), and every run acquires the slot
+  through `acquireCredentialSlot` — a process-wide FIFO mutex that writes the
+  target account's verified credential into the slot, then restores the
+  primary's on release. `SlotUnavailableError` aborts a run instead of letting
+  it silently spend the wrong account when the primary cannot be verified.
+  Remaining boundary: the mutex is **per process**, so two DSH hosts (web +
+  headless) running agy at the same moment can still interleave slot writes;
+  keep `maxConcurrent` at 1 on Windows and avoid driving two hosts at once.
+  A cleaner upstream fix would be per-HOME keyring namespacing in agy.
+
 ## Not forwarded to agy
 
 - **Images as true multimodal blocks** - agy print mode (`-p`) has no image

package/docs/README.en.md

@@ -52,6 +52,7 @@ In DSH, click the `AGY (n)` badge and choose **Add Account** to complete Google
 | `defaultModel` | `DSH_AGY_DEFAULT_MODEL` | agy default | Default model slug |
 | `defaultEffort` | `DSH_AGY_DEFAULT_EFFORT` | model default | `low`, `medium`, or `high` |
 | `timeoutMs` | `DSH_AGY_TIMEOUT_MS` | `600000` | Activity watchdog timeout in ms |
+| `longToolIdleMinutes` | `DSH_AGY_LONG_TOOL_IDLE_MINUTES` | `240` | Idle budget (minutes) while an agy tool step is still open; `0` = keep `timeoutMs` so a wedged tool fails fast |
 | `workspaceRoot` | `DSH_AGY_WORKSPACE_ROOT` | session cwd | Workspace root |
 
 ## `/agy` commands

package/docs/README.es.md

@@ -52,6 +52,7 @@ En DSH, haga clic en la insignia `AGY (n)` y seleccione **Add Account** para com
 | `defaultModel` | `DSH_AGY_DEFAULT_MODEL` | predeterminado de agy | Slug del modelo predeterminado |
 | `defaultEffort` | `DSH_AGY_DEFAULT_EFFORT` | predeterminado del modelo | `low`, `medium` o `high` |
 | `timeoutMs` | `DSH_AGY_TIMEOUT_MS` | `600000` | Timeout del watchdog en ms |
+| `longToolIdleMinutes` | `DSH_AGY_LONG_TOOL_IDLE_MINUTES` | `240` | Presupuesto de inactividad (minutos) mientras un paso de herramienta sigue abierto; `0` = mantener `timeoutMs` |
 | `workspaceRoot` | `DSH_AGY_WORKSPACE_ROOT` | cwd de la sesión | Raíz del workspace |
 
 ## Comandos `/agy`

package/docs/README.pt-BR.md

@@ -52,6 +52,7 @@ No DSH, clique no selo `AGY (n)` e selecione **Add Account** para concluir o log
 | `defaultModel` | `DSH_AGY_DEFAULT_MODEL` | padrão do agy | Slug do modelo padrão |
 | `defaultEffort` | `DSH_AGY_DEFAULT_EFFORT` | padrão do modelo | `low`, `medium` ou `high` |
 | `timeoutMs` | `DSH_AGY_TIMEOUT_MS` | `600000` | Timeout do watchdog, em ms |
+| `longToolIdleMinutes` | `DSH_AGY_LONG_TOOL_IDLE_MINUTES` | `240` | Orçamento de inatividade (minutos) enquanto um passo de ferramenta está aberto; `0` = manter `timeoutMs` |
 | `workspaceRoot` | `DSH_AGY_WORKSPACE_ROOT` | cwd da sessão | Raiz do workspace |
 
 ## Comandos `/agy`

package/docs/adr-012-account-pool.md

@@ -37,12 +37,16 @@ Spawning agy for Account $X$ simply injects `HOME=/path/to/~/.dsh/agy-accounts/a
 ### 2. Sticky Sequential Drain Strategy (按模型家族顺次耗尽)
 - **Family-Scoped Cooldown**: Quota is tracked per model family (`google`, `anthropic`, `openai`). If Account A hits 429 on Claude 4.6, only Account A's `anthropic` family is marked in cooldown. Account A can still serve Gemini 3.7 Flash requests.
 - **Sticky Affinity**: Requests stick to the first healthy account (e.g., Account A) to maximize conversation continuity and token caching.
-- **Transparent Fallback**: When an in-flight request on Account A receives 429 / quota exceeded, the adapter records cooldown on Account A for that family and immediately transparently retries on Account B in the same streaming span. The user experiences zero interruption.
+- **Transparent Fallback**: When a request on Account A is refused for an account-scoped reason (429 / quota, region eligibility, transport-flavoured auth, soft capacity), the adapter records the cooldown or rotates the sticky pointer, and the SAME step continues on Account B — either through DSH's automatic retry (the first-span case) or by discarding the dead recording and re-assembling on the next account (the post-tool case). The user experiences no interruption and never resends.
+  - Account failover is bounded by the pool: DSH's retry budget is `min(enabled accounts - 1, 2)` retries, and the post-tool failover allows at most `enabled accounts - 1` re-spawns per session inside a 2-minute window. A deterministic (request-level) failure is never failed over, so a broken request cannot loop through the pool re-running its tools.
+  - A genuine sign-out (`AUTH`) is retryable only when the pool has another account that may still be signed in; the failed account is flagged `authRequired` and leaves rotation.
 
 ### 3. Cooldown & Reset Handling
 - If the error contains a server reset time (or `Retry-After`), cooldown is set to `Date.now() + resetMs`.
-- Default cooldown window is 10 minutes (tiered up to 60 minutes on consecutive 429s).
+- Default cooldown window is 15 minutes (`defaultCooldownMs`, tiered up to 60 minutes on consecutive 429s).
 - Cooldown expires automatically; when Account A recovers, it resumes primary position.
+- Soft capacity refusals ("model overloaded" / high traffic) NEVER set a cooldown: nothing was spent. They rotate the pointer and are retryable anyway, so an idle account can still serve the turn.
+- "All accounts unavailable" reports the real state (N in cooldown, N signed out) with the earliest availability, which includes quota-window resets and not just cooldowns; a reset that is ≤ 20s away is waited out instead of failing the turn.
 
 ### 4. Optional Per-Account Proxy (防关联支持)
 Each account can optionally specify a dedicated proxy URL (`socks5://...` or `http://...`). When set, `ALL_PROXY` / `HTTPS_PROXY` / `HTTP_PROXY` are scoped exclusively to that account's child process. If unconfigured, it inherits the global host environment proxy.

package/docs/merge-2026-10-13-review.md

@@ -0,0 +1,102 @@
+# 2026-10-13 两线合并评审清单
+
+本次会话把「上游 main 线」与「被舍弃的工作区快照」两条分叉合并为一个可构建、可测试的状态。
+本文是一次性评审依据：逐项列出**做了什么取舍、为什么、怎么改回去**。
+
+## 1. 背景与根因
+
+1. 你记忆里「已经修好」的 Windows 凭据隔离 + 客户端增强，**从未提交、从未推送**——它只存在于
+   `D:\proj\dsh-agy-link` 的未提交工作区里（v0.4.44 草案，2026-10-03 起草）。
+2. `origin/main`（v0.4.46）**只包含那条警告横幅的披露逻辑**，没有任何真实隔离实现：
+   - `src/client/locales.ts` 的 `pool.winIsolationWarn` 文案
+   - `src/index.ts` 的 `credentialIsolationLimited` 状态字段
+   - `src/host/adapter.ts` 的 `warnOnce` 日志
+   - 全仓库没有 `src/host/wincred.ts` / `src/host/credential-slot.ts`
+3. 运行中的 GUI 通过 **Junction** 直接吃本目录的 `dist/`：
+   `C:\Users\zand\.dsh\profiles\desktop\node_modules\dsh-agy-link -> D:\proj\dsh-agy-link`。
+   清理工作区 + `pnpm install` 触发 `prepare: npm run build` 后，`dist/` 由**上游源码**重建，
+   于是界面从「你的版本」切成了「上游版本」：横幅出现、样式回退。**源码一份没丢**。
+
+## 2. 三方对象
+
+| 角色 | 引用 | 内容 |
+| --- | --- | --- |
+| 基线（第一父） | `main` = `b29e29b` | 上游 0.4.46（含 20 个提交） |
+| 你的工作区快照（第二父） | `recovered/local-afternoon` = `b26271e` | 0.4.44 草案，+9452 / −80 |
+| 合并结果 | `merge/restore-afternoon` | `140d257` 起的合并提交 |
+
+相对 `main` 的净变更：**28 个文件、+3401 / −76**；新增 11 个文件：
+
+```
+.npmrc                              scripts/repair-pool.mts
+scripts/verify-credential-store.mts src/client/decisions.ts
+src/client/images.ts                src/host/credential-slot.ts
+src/host/media-route.ts             src/host/wincred.ts
+test/client-enhancers.test.ts       test/credential-slot.test.ts
+test/media-route.test.ts
+```
+
+## 3. 十处冲突的落定
+
+| # | 文件 | 处理 | 理由 | 想改回去 |
+| --- | --- | --- | --- | --- |
+| 1 | `CHANGELOG.md` | 两边全保留；你的 0.4.44 草案提为顶栏 **0.4.47** 并加合并说明 | 历史不丢、本次发布有说明 | 直接编辑 |
+| 2 | `package.json` | 取 main（**不加 `mermaid` 依赖**），版本 bump `0.4.47` | 上游改为内置 `assets/mermaid.min.js`，不需要 npm 依赖 | 加回依赖 |
+| 3 | `assets/mermaid.min.js` | 取 main（3.57 MB 官方构建） | 与 main 的 `/plugins/agy-link/mermaid.min.js` 路由匹配 | 换回你的 4.37 MB 版 |
+| 4 | `src/client/mermaid.ts` | 取 main（337 行，无 CDN 回退 + 错误炸弹清理 + `ctx.effect` 生命周期） | 你的 598 行版本已被上游同类修复覆盖 | 恢复快照文件 |
+| 5 | `src/client/toolview.ts` | main 的 i18n `tv.delete` + 你的 `question` / `image` 卡片类型 | 两边都是增量，可兼得 | 直接编辑 |
+| 6 | `src/client/index.ts` | main 的 `ctx.effect` 安装器 + 你的图片/决策增强（**已补真实 disposer**） | 保留 main 的重载不叠加 observer 契约 | 直接编辑 |
+| 7 | `src/host/quota.ts` | main 的**异步** Windows 凭据读取 + 你的「持久存储优先」主账号解析；刷新写盘保留你的保形 `doc` 并补 main 的 `0600` | 你那份 `execFileSync` 会让事件循环冻结（上游专门修过）；你的持久库逻辑是防串号关键 | 直接编辑 |
+| 8 | `src/host/runner.ts` | 取 main（`nodeScriptEnv` / `ELECTRON_RUN_AS_NODE`） | 同上问题上游已有更完整修复 | 直接编辑 |
+| 9 | `test/mapper.test.ts` | **两边用例全保留**（#43 思考横幅 + 你的图片重写） | 两者互不冲突 | 无需 |
+| 10 | `test/proxy-keychain.test.ts` | 取 main | 与 main 的 kill 升级语义一致 | 恢复快照版 |
+
+## 4. 语义变更（你已确认）
+
+1. **删除 `credentialIsolationLimited` 警告**：状态字段、adapter `warnOnce`、设置面板横幅、
+   四语言文案全部移除。理由：凭据槽交接恢复后轮巡是真隔离，横幅已是假陈述。
+   `docs/KNOWN-GAPS.md` 改写为「已实现交接 + 剩余边界」。
+2. **`node:test` 下跳过真实凭据槽**（检测 `NODE_TEST_CONTEXT`）：你的严格拒绝语义与上游
+   「额度耗尽自动转移账号」测试在 Windows 上互斥；若不跳过，`pnpm test` 会往**真实凭据管理器**
+   写测试凭据。生产的严格语义（`SlotUnavailableError`：无法验证主账号 / 副账号无凭据 /
+   同一授权）**完整保留**；`test/credential-slot.test.ts` 仍通过注入 IO 直测交接逻辑。
+
+## 5. 被取代或删除的内容（知情项）
+
+| 内容 | 去向 | 说明 |
+| --- | --- | --- |
+| `src/client/mermaid.ts`（你的 598 行） | 未进入合并 | 上游同类修复已覆盖（`MERMAID_CSS` 抑制 + `cleanupTempElements` 清扫） |
+| `test/client-enhancers.test.ts` 内 5 个 mermaid 用例 | 已删除 | 随上一条；保留图片相关用例 |
+| `quota.ts` 中经 `wincred` 的同步 CredRead | 改用 main 异步实现 | `wincred.ts` 仍用于槽交接与系统代理探测 |
+| 4 条 `pool.winIsolationWarn` 文案 | 已删除 | 横幅移除后成为死数据 |
+
+## 6. 保留清单
+
+**来自 main（上游）**：适配器 spawn 窗口槽位归属 try/finally、异步钥匙串/凭据管理器读取、
+桌面端思考可见性 #43、内置 mermaid 渲染器（无 CDN）、模型可见性开关、`prepare` 构建、
+账号池/数据加固、`docs/KNOWN-GAPS.md` 框架。
+
+**来自你的快照**：`credential-slot.ts`（持久主凭据库 + 身份校验 + 进程级 FIFO 互斥 +
+`SlotUnavailableError`）、`wincred.ts`（CredRead/CredWrite/CredDelete + WinINet 代理探测）、
+`media-route.ts` + `images.ts` + `decisions.ts`（本地图片路由 / 灯箱 / 决策芯片）、
+`repair-pool.mts`、`verify-credential-store.mts`、`.npmrc`、`.gitignore` 的 `Microsoft/`、
+以及 4 个测试文件（`credential-slot` / `media-route` / `client-enhancers` / `mapper` 图片用例）。
+
+## 7. 验证
+
+```
+pnpm run check   -> exit 0（0 错误）
+pnpm run build   -> exit 0（dist/client.js 187KB、dist/index.js 1.32MB、dist/assets/）
+pnpm test        -> exit 0（31 个测试文件、328 用例通过、0 失败）
+```
+
+`dist/` 已由合并后的源码重建。
+
+## 8. 生效方式与后续
+
+- **客户端半**：刷新页面即生效——`dist/client.js` 中已无 `win-isolation` 渲染，横幅不会再出现。
+- **宿主半**：需要**重启 DSH / 重载插件**。当前 `GET /plugins/agy-link/status` 仍返回
+  `credentialIsolationLimited`，证明运行中的宿主进程还是旧代码；宿主半决定凭据交接是否真的生效。
+- **未做**：没有 `push`、没有打 tag、没有发布（遵循 AGENTS.md：发布需你显式下令）。
+- **已知边界**：凭据槽互斥锁是**进程级**的——web 与 headless 同时驱动 agy 仍有交叉写风险；
+  Windows 上建议 `maxConcurrent=1`，并避免两个宿主同时跑。彻底解决需上游 agy 按 HOME 命名 keyring。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-agy-link",
-  "version": "0.4.44",
+  "version": "0.4.49",
   "description": "Google Antigravity (agy CLI) models for DeepSeek Harness — stream Gemini/Claude/GPT-OSS subscriptions into DSH with thinking, tool activity, token usage and in-GUI Google OAuth login.",
   "type": "module",
   "license": "MIT",