tianshu-mcp 0.5.4 → 0.5.6

package/CHANGELOG.en.md

@@ -8,11 +8,45 @@ Chinese version: [CHANGELOG.md](CHANGELOG.md)
 
 ---
 
-## [Unreleased]
+## [0.5.6] - 2026-09-22
+
+### Added
+
+- Add the Qoder CN GUI adapter: installation discovery, full-path workspace binding and import, default/custom model selection, and persisted reasoning settings verified through Model Management.
+- Integrate dispatch, liveness, objective acceptance, and same-session rework. Automatic and manual rework write a plan before sending its filename, full path, and complete text.
+- Report model source and actual settings; retain permission mode and disclose global reasoning preferences. Approvals require the user, questions use dedicated controls, and uncertain submissions are never repeated automatically.
+- Windows default/custom model, new workspace and same-session repair acceptance passed on a real desktop. macOS remains research and dispatch is disabled. See the [Qoder guide](docs/qoder-cdp.en.md).
+- Isolate unit test files to prevent discovery command mocks from contaminating Git baseline tests; exclude temporary probes from lint.
+
+### Tests
+
+- Full suite: **826 passed / 12 skipped** (Windows 10 x64, Node 24.18.0; 78 test files passed plus 3 real-browser files skipped by design), roughly 60 cases more than v0.5.5: discovery priority (explicit → D drive → registry/shortcuts → standard directories), wrong installation paths and identity checks, CJK/space/same-name paths, workspace import and read-back failure, cross-group model ambiguity and unsupported tiers, a save that did not persist, uncertain sends and answer submissions never resent, this-turn-bound completion judging (old replies and a static screen do not qualify), same-session repair, unconfirmed cancellation/timeout stops, and macOS branches failing closed.
+- Real-browser gate cases (`TIANSHU_VISUAL_BROWSER_TEST=1`) remain 12/12 on this Windows 10 machine; `npm pack` content validation plus a clean-consumer install and strict stdio check pass locally.
+- Uncovered items are stated plainly: Qoder cancellation, question answering and login/quota/network waiting classification are **covered by hermetic integration tests only**, and macOS has no real GUI validation. See [HANDOFF §9.12](HANDOFF.md).
+
+### Planned
+
+- The Qoder CN macOS hardware matrix (stays `research`; dispatch disabled until verified).
+- Hardware verification of Qoder CN cancellation, question answering and login/quota/network waiting classification (currently covered by hermetic integration tests only).
+
+## [0.5.5] - 2026-09-20
+
+### Added
+
+- **Added the Kimi Code GUI adapter (`agentId=kimicode`, the fourth GUI agent)**: the Kimi Code desktop app (Moonshot AI, measured 1.0.2) is a **plain Electron install** — injecting `--remote-debugging-port` and driving it over CDP is enough, with **no** MSIX COM activation.
+  - **Two-renderer CDP driving**: the model menu, thinking tiers and execution-mode menu are rendered by the app's `browserOverlayOpenMenu()` into a separate `Kimi Browser Overlay` renderer process (measured: after clicking `model-pill` the main window's DOM node count is unchanged and no menu node appears), while the workspace menu and the "switch model" dialog stay in the main window → the client holds both pages and excludes the `Screenshot` target.
+  - **Workspace binding and import**: the **normalized full path** is the sole criterion (same-name different-directory always fails closed, never guesses an entry); an unregistered workspace is imported through the native "add workspace" dialog (`#32770`; Win32 coordinate clicks plus `WM_SETTEXT`/`WM_GETTEXT`), and after binding the panel selection and the `ws-chip` text are read back.
+  - **Three-stage model selection**: pill read-back → direct pick in the overlay shortcut menu → "more models…" → search and exact row pick in the main window's "switch model" dialog (the only entry for unofficial models). Thinking tiers are validated against **the tier set the UI actually renders** (official `Low/High/Max`, unofficial `On/Off`); requesting a tier the UI does not render fails loudly with `model_mismatch` before sending and is never silently kept.
+  - **Execution mode is forced to "fully automatic"** and read back after switching; the `run_task.reasoningLevel` domain grows to `low/medium/high` plus `max/on/off`.
+  - **Run detection**: `button.stop` (`aria-label="中断"`) and `button.send.is-starting` are the authoritative run signals; sending uses a marker plus a bounded 60s confirmation (the session id and the landed user message are required anchors) and is **never re-sent**.
+  - **Six `needs_user` kinds with `continue_task` recovery**: `close_existing_instance` / `login_required` / `user_confirmation` / `agent_question` / `system_permission` / `setup_recovery`. Answering a question writes back to the recorded session (without re-sending the brief), a user confirmation only re-attaches for observation, and environment kinds re-send the full brief; when the recorded session cannot be located the run fails closed with `session_lost`.
+  - **Cancellation and dispatch guard**: following the Codex M14 semantics, `cancel_task` clicks `button.stop` best-effort and bounded-waits (`gui.cancelWaitMs`) for the UI to go idle, stating plainly when the stop is unconfirmed; before dispatching, a still-running managed instance is stopped best-effort, and if it never goes idle the dispatch is rejected with `instance_busy`.
+  - **Machine verification (Windows 10 x64 + Kimi Code 1.0.2)**: the success path, unregistered-workspace import with auto-acceptance, and the failure → rework → re-acceptance **same-session loop** all pass. **Cancellation, question answering and same-name workspace ambiguity are covered by hermetic integration tests only** (no hardware stop click, no real question card triggered); macOS is `research` and fail-closed (executable discovery and native-dialog driving are unmeasured on macOS).
 
 ### Planned
 
 - More external AI-Agent adapters (a new agent = one profile + an optional adapter file).
+- The Kimi Code macOS hardware matrix, plus hardware verification of cancellation / question answering / same-name ambiguity (currently covered by hermetic integration tests only).
 - TraeWork executable discovery and native-dialog driving on macOS (currently fail-closed).
 - Optional project-level skill seeding (by default nothing is written into target repos).
 - Cancel/rework/new-project matrices for the Codex and ZCode GUI drivers on macOS (both remain `research` on darwin).
@@ -22,6 +56,10 @@ Chinese version: [CHANGELOG.md](CHANGELOG.md)
 - ZCode's **automatic import of an unregistered project** cannot complete on Windows: the native-panel script relies on `SetForegroundWindow` to bring the dialog forward before activating its address bar, but a child process of a background MCP server is refused by Windows, so the address-bar Edit never appears and the script spins until its deadline (measured: 56s, then classified by `budget.check()` as an exhausted setup budget). PowerShell's stdout is also block-buffered through a pipe, so killing the process loses the buffer and not a single `native:` stage reaches the log, misdirecting diagnosis. Workaround: add the target directory to the ZCode project list manually first; the fix direction is to grab the foreground with `AttachThreadInput` inside the script, or to use a supported ZCode registration entry point.
 - Cross-round verdict-flip circuit breaking for AI content validation (this round covers it with caching plus sampling; reconsider if hardware data still shows churn), cross-task cache sharing, and reference-image/design-diff comparison.
 
+### Fixed
+
+- **Completed the two contract-mapping assertions issue #13's plan §5 G requires (added after the v0.5.4 release)**: `CONTENT_TIMEOUT` had its classification logic implemented but tests only asserted the transport-level `outcome.timeout`, and the stub's `sleep` mode was never exercised by any case. `visual-content-command` now has an end-to-end assertion (a timeout yields a single `blocked` item with `CONTENT_TIMEOUT`, still warning-only, with temp input files removed), plus a success-path temp-input-removal assertion and a `visual doctor` assertion for the "multi-rule total budget exceeds `roundTimeoutMs` → advisory only" branch. The v0.5.4 tag (`ea797d1`) shipped 644 cases; master now has 647.
+
 ---
 
 ## [0.5.4] — 2026-09-16
@@ -776,7 +814,13 @@ project → pick model and reasoning level → send instructions → run detecti
 
 ---
 
-[Unreleased]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.4.1...HEAD
+[0.5.6]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.5...v0.5.6
+[0.5.5]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.4...v0.5.5
+[0.5.4]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.3...v0.5.4
+[0.5.3]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.2...v0.5.3
+[0.5.2]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.1...v0.5.2
+[0.5.1]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.0...v0.5.1
+[0.5.0]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.4.1...v0.5.0
 [0.4.1]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.4.0...v0.4.1
 [0.4.0]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.3.4...v0.4.0
 [0.3.4]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.3.3...v0.3.4

package/CHANGELOG.md

@@ -7,11 +7,45 @@
 
 ---
 
-## [未发布]
+## [0.5.6] - 2026-09-22
+
+### 新增
+
+- 新增 Qoder CN GUI 适配：安装发现、完整路径工作区绑定与导入、默认/自定义模型选择、模型管理思考等级保存回读。
+- 接入已有 MCP 开发、运行检测、客观验收和原会话返修流程；自动与手动返修均先保存修复计划，再发送文件名、完整路径与全文。
+- 新增模型来源与实际设置报告；保留权限模式，披露全局思考偏好影响；审批不代批，提问使用专用答题控件，提交不明不重发。
+- Windows 默认/自定义模型、新工作区与原会话返修真机验收通过；macOS 保持 research 并禁止派发。详见 [Qoder 操作说明](docs/qoder-cdp.md)。
+- 恢复单元测试文件隔离，修复安装探测命令 mock 泄漏导致 Git 基线测试依赖执行顺序的问题；lint 排除临时探测目录。
+
+### 测试
+
+- 全量 **826 passed / 12 skipped**（Windows 10 x64，Node 24.18.0；78 个测试文件通过 + 3 个真实浏览器文件按设计 skip），较 v0.5.5 净增约 60 项：安装发现优先级（显式 → D 盘 → 注册表/快捷方式 → 标准目录）、错误安装路径与身份校验、中文/空格/同名路径、工作区导入与回读失败、模型跨组重名与档位不支持、保存未生效回读、发送/答题提交不明不重发、本轮绑定完成判定（旧回复与静止不触发）、原会话返修、取消与超时停止不确定、macOS 平台分支 fail-closed。
+- 真实浏览器门禁用例（`TIANSHU_VISUAL_BROWSER_TEST=1`）在 Windows 10 本机维持 12/12；`npm pack` 内容校验与干净消费者安装 + 严格 stdio 检查在本机通过。
+- 未覆盖项如实标注：Qoder 的取消真停、提问续答与登录/额度/网络等待分类**仅由 hermetic 集成测试覆盖**；macOS 未做真机 GUI 验证。详见 [HANDOFF §9.12](HANDOFF.md)。
+
+### 计划中
+
+- Qoder CN 的 macOS 真机验证矩阵（保持 `research`，未验证前禁止派发）。
+- Qoder CN 取消真停 GUI、提问续答与登录/额度/网络等待分类的真机验证（当前仅 hermetic 集成测试覆盖）。
+
+## [0.5.5] - 2026-09-20
+
+### 新增
+
+- **新增 Kimi Code GUI 适配（`agentId=kimicode`，第四个 GUI agent）**：Kimi Code 桌面端（Moonshot AI，实测 1.0.2）是**普通 Electron 安装**，以 `--remote-debugging-port` 注入后经 CDP 驱动，**不需要** MSIX COM 激活。
+  - **双渲染进程 CDP 驱动**：模型菜单 / 思考档位 / 执行模式菜单经应用内 `browserOverlayOpenMenu()` 渲染在独立的 `Kimi Browser Overlay` 渲染进程（实测点击 `model-pill` 后主窗口 DOM 节点数不变、不产生任何菜单节点），工作区菜单与「切换模型」对话框仍在主窗口 → 客户端同时持有两个页面，并排除 `Screenshot` target。
+  - **工作区绑定与导入**：以**归一化完整路径**为唯一判据（同名不同目录一律 fail-closed，绝不猜一个点），未登记的工作区经原生「添加工作区」对话框（`#32770`；Win32 坐标点击 + `WM_SETTEXT`/`WM_GETTEXT`）导入，绑定后回读面板选中项与 `ws-chip` 文本。
+  - **模型三级选择**：pill 回读 → overlay 快捷菜单直选 → 「更多模型…」→ 主窗口「切换模型」对话框搜索精确选行（非官方模型的唯一入口）；思考档位按**界面实际渲染的档位集合**校验（官方 `Low/High/Max`，非官方 `On/Off`），请求界面不存在的档位在发送前以 `model_mismatch` 响亮失败，绝不静默沿用。
+  - **执行模式强制「完全自动」**并在切换后回读确认；`run_task.reasoningLevel` 取值域扩展为 `low/medium/high` + `max/on/off`。
+  - **运行检测**：`button.stop`（`aria-label="中断"`）与 `button.send.is-starting` 为权威运行信号；发送以标记 + 60s 有界确认（会话 id 与用户消息落地为必需锚点），**绝不重发**。
+  - **`needs_user` 六类与 `continue_task` 恢复**：`close_existing_instance` / `login_required` / `user_confirmation` / `agent_question` / `system_permission` / `setup_recovery`；提问续答写回原会话（不重发任务书）、用户确认仅重连观察、环境类补发完整任务书；定位不到原会话一律 `session_lost` fail-closed。
+  - **取消与重派护栏**：照 Codex M14 语义，`cancel_task` 尽力点 `button.stop` 并在 `gui.cancelWaitMs` 内有界等待界面空闲，未确认停止时终态如实明示；派发前发现未停止的运行先尽力停止，仍不空闲以 `instance_busy` 拒绝。
+  - **真机验证（Windows 10 x64 + Kimi Code 1.0.2）**：成功路径、未登记工作区导入 + 自动验收、失败 → 返修 → 再验收**同会话闭环**均已通过。**取消、提问续答、同名工作区歧义仅由 hermetic 集成测试覆盖**（未在真机点停、未触发真实提问卡片）；macOS 为 `research` 且 fail-closed（可执行探测与原生对话框驱动未在 macOS 实测）。
 
 ### 计划中
 
 - 更多外部 AI-Agent 适配（新 agent = 一个 profile +（如需）一个 adapter 文件）。
+- Kimi Code 的 macOS 真机验证矩阵，以及取消 / 提问续答 / 同名歧义的真机验证（当前仅 hermetic 集成测试覆盖）。
 - TraeWork 在 macOS 下的可执行探测与原生对话框驱动（当前 macOS 分支 fail-closed）。
 - 可选的项目级技能播种（默认不写入目标项目仓库）。
 - Codex 与 ZCode GUI 的 macOS 取消/返修/新建项目矩阵（当前两者 darwin 均保持 `research`）。
@@ -20,6 +54,10 @@
 - ZCode **未登记项目**的自动导入在 Windows 上无法完成：原生面板脚本靠 `SetForegroundWindow` 抢前台来激活地址栏，而后台 MCP server 的子进程会被 Windows 拒绝，地址栏 Edit 永不出现，脚本空转到 deadline（实测 56s 后由 `budget.check()` 归类为 setup 预算耗尽）；且 PowerShell 的 stdout 在管道里被缓冲、进程被 kill 后缓冲丢失，日志里连一条 `native:` 阶段都看不到，排障方向被误导。临时对策：先在 ZCode 中手动把目标目录加入项目列表；修复方向是脚本内改用 `AttachThreadInput` 抢前台（或改走 ZCode 受支持的登记入口）。
 - AI 内容校验的跨轮判定翻转熔断（本轮以缓存 + 采样覆盖；若真机数据显示仍扰动再议）、跨任务缓存共享、参考图/设计稿差异比对。
 
+### 修复
+
+- **补齐 issue #13 计划 §5 G 要求的两项契约映射断言（v0.5.4 发布后补）**：`CONTENT_TIMEOUT` 此前只实现了分类逻辑、测试仅断言传输层 `outcome.timeout`，判定桩的 `sleep` 模式从未被用例使用；现补 `visual-content-command` 的端到端断言（超时 → 单项 `blocked` + `CONTENT_TIMEOUT` + 仍为仅告警 + 临时输入文件已删除），并补成功路径的临时输入删除断言，以及 `visual doctor` 的「多规则总预算超 `roundTimeoutMs` 只给建议值」分支断言。v0.5.4 tag（`ea797d1`）的用例数为 644，master 现为 647。
+
 ---
 
 ## [0.5.4] — 2026-09-16
@@ -671,7 +709,13 @@ Codex 桌面端改为 **GUI 驱动**：新增 `codex-gui` adapter，通过 MSIX
 
 ---
 
-[未发布]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.4.1...HEAD
+[0.5.6]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.5...v0.5.6
+[0.5.5]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.4...v0.5.5
+[0.5.4]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.3...v0.5.4
+[0.5.3]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.2...v0.5.3
+[0.5.2]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.1...v0.5.2
+[0.5.1]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.5.0...v0.5.1
+[0.5.0]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.4.1...v0.5.0
 [0.4.1]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.4.0...v0.4.1
 [0.4.0]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.3.4...v0.4.0
 [0.3.4]: https://github.com/lanlan0811/tianshu-mcp/compare/v0.3.3...v0.3.4