@suifei/dsh-ai-team 0.0.0-stage → 0.3.3

package/CHANGELOG.md

@@ -0,0 +1,98 @@
+# Changelog
+
+## 0.3.3 — 2026-11-09
+
+- 产品显示名改为「文渊阁」（Wenyuan Pavilion），取藏修、校勘、纂辑之意；
+- 保持包名 `@suifei/dsh-ai-team` 与内部 Team 域对象不变。
+
+## 0.3.2 — 2026-11-09
+
+- 产品显示名改为「三省」（Three Departments），副标「三省六部」；
+- 修复热加载时 permission 服务尚未就绪导致 `BOOT_FAILED / permission services are unavailable`，Host 显式 inject `permissionPresets` 与 `sandboxPolicy`；
+- 四象限空 items / null taskNumber / 无 taskId 的决策项做健壮处理，避免点击异常与 `#null` 文案。
+
+## 0.3.1 — 2026-11-09
+
+- SkillOpt 完全内置：7 个 `skillopt_*` 工具、Python engine bootstrap、`skillopt-sleep` 技能随插件分发，不再装到用户 DSH；
+- 冷恢复重新 mount member preset，修复重启后 `pwsh/read/write` 工具消失；
+- 失败决策自动带 taskId；跨 Project 收件箱可一键打开 Task；
+- 双真实 provider SSE：DeepSeek 与 Zhipu 各自项目独立分桶；
+- 测试 82 条。
+
+## 0.3.0 — 2026-11-09
+
+- 并发 ProjectRuntime：启动打开全部 Project，各自 ledger/Waker/SSE/Team 工具路由；
+- 跨 Project 收件箱：聚合、项目标记、按项目 ack、打开 Task 前切项目；
+- V-10：新增真实 turn 中途降级 E2E 测试钩子（生产 403），守卫在工具调用前修复权限；
+- billing/网络：测试专用 `ai-team-fault` LLM adapter，真实 Agent loop 验证 402 与 ECONNRESET 分类/恢复；
+- 技能：新增 `ai-team-files` provider，发现 `~/.dsh/skills` 与项目技能；nuwa/distilly/darwin/SkillOpt 四个技能安装并由真实成员加载。
+
+## 0.2.6 — 2026-11-09
+
+- 工具调用前权限守卫：managed member 若在 turn 中途被降级，下一步工具前先修复为 danger-full-access；修不好则拒绝；
+- 非成员工具调用不受影响；
+- V-10 从“仅激活时检查”前移到每次工具调用。
+
+## 0.2.5 — 2026-11-09
+
+- 成员 config 新增 `preset`：`standard / ptc / minimal / cordis`，默认 `standard`；
+- 资料抽屉可选择运行配方；
+- 激活时 mount 对应 preset，未知配方被校验拒绝。
+
+## 0.2.4 — 2026-11-09
+
+- 修复成员工具链：`agents.create` 的 setup 里 mount `standard` preset，成员获得 pwsh/read/write/edit/skill 等完整工具；
+- 真实成员用 pwsh 写文件、用 skill 加载 DSH 技能并写结果；
+- 健康检查暴露 `presets` 与 `skillsService`。
+
+## 0.2.3 — 2026-11-09
+
+- 成员 activation/wake 注入 DSH `skills` 目录摘要（名称/描述/何时用，最多 12 条）；
+- 完整技能正文由成员自己的 `skill` 工具按需加载；
+- 保留 `<dataDir>/skills/<name>/SKILL.md` 有界注入。
+
+## 0.2.2 — 2026-11-09
+
+- 多 Project 单活动切换：每 Project 独立 ledger，`projects.json` 索引；
+- Host `project.create` / `project.activate`；
+- UI 新建项目弹窗 + 项目列表切换；
+- 切换后重拉 bootstrap、重连 SSE，未完成 Task 由 `team/startup` 恢复；
+- 已知限制：同一时刻只开一个 Project runtime，非活动项目不并发唤醒。
+
+## 0.2.1 — 2026-11-09
+
+- QA checklist 硬门禁：`qaRequired` + `qaChecklist`，全部通过才能 accept。
+- 启动恢复：`team/startup` 恢复未完成 Task，无定时器；真实重启 E2E 跑通。
+- 抽出 `ProjectRuntime`（默认项目行为不变），为多 Project 做准备。
+- 修复资料抽屉白屏（early-return 前引用 selectedTask / data null）。
+- 真实浏览器验证 DM：`member.message` + wake，成员 session 前进。
+
+## 0.2.0 — 2026-11-09
+
+M2.5「团队进化」迭代。
+
+### 新增
+
+- 成员资料：name / role / personality / profileVersion；Human 可编辑，空个性按“不修改”处理。
+- 自我进化：`team.evolve` 只允许 Agent 改自己的 personality；验收后 reflection wake 触发；真实跑通并写 `member/distilled`。
+- 模型蒸馏：`member.distill(mode:'model')` 由 Host 单步调用 `llm.stream`，无工具、固定 JSON schema；真实跑通。
+- 人才库：`member.leave` / `member.rejoin`，身份与 session 保留；`member.takeover` 接遗留 Task，保留 claim/takeover 历史。
+- QA：`qaRequired` 硬门禁、`qaChecklist` 逐项通过、QA 优先自动 reviewer；真实 QA 自动验收。
+- 备忘/技能：`team.note` / `team.skill` / `team.recall`；`<dataDir>/skills/<name>/SKILL.md` 有界加载进 prompt。
+- 私信：`member.message` 只唤醒目标成员。
+- 关注：`task.follow`；验收后给仍关注者生成 `followed_task_update`。
+- 收件箱：真实 `data.inbox`、ack、打开 Task；失败恢复 UI（改配置 + retry）。
+- 启动恢复：`team/startup` 恢复未完成 Task，无定时器；真实重启 E2E。
+- 任务动作 UI：取消/重开/删除/改派；Thread 回复/验收/退回真实命令。
+- 安装/卸载/升级文档与 profile 隔离检查。
+
+### 测试
+
+- `node --test` 71/71。
+
+### 已知边界
+
+- 单机、单 Human；多 Project 仍以默认 Project 为主。
+- 技能文件只进 prompt，不是可执行工具注册。
+- 外部失败注入仅覆盖真实 401；billing/网络注入仍待。
+- 多 Project、V-10、真实卸载/升级演练未完成。

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 suifei
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -1,3 +1,39 @@
-# Temporary Holding Version
+# 文渊阁 · @suifei/dsh-ai-team
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+> 文渊阁：藏修、校勘、纂辑。AI Team mode for DeepSeek Harness:
+
+AI Team mode for DeepSeek Harness: **projects, owned tasks, threads, inbox and an append-only team ledger** — one Human sentence, a team of named persistent agents keeps going, and only direction / trust / irreversible steps come back.
+
+> M1/M2 status: the package mounts a **three-pane workbench** (left projects + agents, center `主对话 / 看板 / 态势`, right thread). The app ledger, task/thread/inbox domain and `TeamService` are implemented and covered by `node --test` (82 tests); `bootstrap`, `POST .../commands`, `GET .../board`, `GET .../quadrants` and the event-driven `GET .../events` SSE are wired to a real default-project ledger. The profile patch enforces `danger-full-access`, and the Host refuses to boot a weaker profile; `member.activate` now creates a real continuable child through the Team Host Agent, binds the session, sets/verifies `danger-full-access`, and records it in the ledger. Real activation and `member.config.update` were verified on the ai-team profile; the Host subscribes to `agent/request-error` for classification/retry/fallback (auth/billing fail closed) and to `session/event` for recovery/reset; the member row shows effective model/context/version/last failure. A member-only `team` tool exposes tasks/inbox/claim/submit/thread/reply/review through `TeamService`; the Waker delivers wake messages and records `wake/delivered`/`wake/failed`. The direct `agents.create + followup` path now completes a real two-Agent loop: 沈舟 claimed Task #1, replied in the Thread and submitted; 林岚 independently accepted it (`reviewerId=a2`, `selfExecuted=false`), leaving the task `done`. A follow-up probe wrote one Human sentence into the Main stream; Lead a1 created Task #1 itself, claimed/executed/submitted it, and a2 independently accepted it. The workbench Composer is wired to the real `stream.post` / `task.create` commands, and the Agents “+” opens an Add Member form that creates, configures, activates and optionally promotes the member to Project owner. A headless-Edge E2E opened 文渊阁, added 沈舟 as owner through the form, sent one sentence, and saw Task #1 appear on the board and in Q1. A second form submission added 林岚3 as reviewer with an explicit accept instruction; the board then showed `Done 2` / `To accept 0`. Host now broadcasts SSE on every ledger append, including member `team` tool calls. M2.5 adds versioned member profiles (name/role/personality), Agent self-evolution via `team.evolve`, a talent pool with leave/rejoin, leftover-task takeover, and QA-first automatic reviewer wake; a real QA member accepted a submitted Task with no manual reviewer instruction. The profile/talent-pool UI was verified in a real browser: saving a new name/personality bumped `profileVersion`, then leaving the team with a successor moved the member to the talent pool and transferred the leftover Task (`taken-over`) to the successor. Acceptance also schedules a reflection wake; a real Agent then called `team.evolve` itself and wrote `member/distilled`, with the new personality visible on the member. Members can add durable notes/skills (`team.note`/`team.skill`); skill bodies are read from `<dataDir>/skills/<name>/SKILL.md` (bounded, traversal-safe) and injected into activation/wake prompts. Activation/wake also injects a bounded DSH `skills` catalog (name/description/when-to-use, max 12) so members can load a full skill through their own `skill` tool. Members now mount an agent preset in `agents.create` setup; a real member used `pwsh` and then `skill` to load `diagnose-windows-sandbox-acl` and wrote the result to a file. The profile drawer can choose `standard / ptc / minimal / cordis` (default `standard`). All four were exercised end-to-end: PTC used `run_code`, minimal and cordis used `pwsh`, standard used `pwsh` + `skill`. `team.recall` searches their own notes/skills/tasks/threads; a Lead can inspect a colleague (`team.member`) and call `team.distill`. Host now has a narrow single-step distillation path: `member.distill(mode:'model')` feeds the member's profile, notes, skills and recent tasks to `llm.stream` with no tools, requires JSON output, then records `member/distilled`. A real run produced a new personality from the old one plus a missed-gate note. Free-form Lead-driven distillation can still drift; the Host path is the product entry. Restart persistence was verified: the same member/session resumed from 63 persisted events (31 zstd frames), and a post-restart Task was claimed and submitted without re-onboarding. The archived-host gate was fixed by not marking members as `origin:'subagent'`. A real 401 also produces `member/llm-failed(auth)`, member `failed`, a Human `decision_required`, and leaves the task pending without retry. See `docs/reports/M2-真实闭环验收.md`. The workbench Thread now sends real `thread.post`, `task.accept` and `task.request_changes`; the task action row also drives `task.cancel`, `task.reopen`, `task.delete` and `task.reassign` (the “demo only” tag is gone). A headless-Edge run accepted a pending Task to `done` from the UI. A Human can DM one member from the profile drawer (`member.message`), which writes to the ledger and wakes only that member; a headless-Edge run sent a DM and verified the member session advanced. The first profile-drawer E2E also exposed and fixed a white-screen bug (`selectedTask` referenced across an early return / null `data`). The Inbox page lists real `data.inbox` decisions and acknowledges them through `inbox.ack`. Follow/unfollow is persisted through `task.follow`; accepting a followed Task creates one `followed_task_update` report for each still-following member. A headless-Edge failure-recovery run used a bad provider to produce a Human decision, acknowledged it, edited the member's provider/maxTokens in the profile drawer, and retried the member back to `running` (`member/recovered`). The latest regression re-ran one sentence → Lead opens a card → QA auto-accept → owner reflection/self-evolution with one wake per stage (20 ledger events, no duplicate human/reflection wakes). On Host startup, unfinished Tasks are woken once via a `team/startup` event (no timers). A real restart test resumed a pending Task: `wake/delivered reason=startup_resume`, then claim → Thread → submit to pending acceptance (15 ledger events; tail repair recorded). The profile drawer now edits provider/model/maxTokens, adds notes/skills, and retries failed members. Task creation can mark `qaRequired` and a `qaChecklist`; non-QA Agents are rejected on accept, the Waker only offers the task to QA, and QA must pass every checklist item before acceptance. M3 adds a test-only `ai-team-fault` LLM adapter for real billing/network failure injection, cross-project Inbox (including opening a task from another project), and a filesystem SkillProvider for installed skills. SSE buckets were verified with two real providers (DeepSeek and Zhipu) producing events in two Projects. Microsoft SkillOpt's seven tools, Python engine and `skillopt-sleep` skill are bundled inside this plugin, not installed into the user's DSH. Version 0.3.3 (文渊阁 / Wenyuan Pavilion; concurrent Projects + cross-project inbox + V-10 guard E2E + billing/network injection + bundled SkillOpt; health reports 0.3.3). Install/uninstall was exercised on a throwaway `ai-team-smoke` profile: health 200 after install, 404 after removal, config clean. — see `CHANGELOG.md`, `docs/05-项目启动与执行计划.md`, `docs/08-异常恢复与模型治理.md` and `docs/reports/M1-验收矩阵.md`.
+
+## Install (dedicated profile)
+
+```sh
+dsh --profile ai-team --from-default-profile web --dump-config   # create the profile from the web template
+dsh plugin --profile ai-team add @suifei/dsh-ai-team             # install from npm
+# or from a checkout:
+# dsh plugin --profile ai-team add E:/works/dsh-ai-team
+dsh --profile ai-team --no-open --port 0                         # boot
+```
+
+The plugin is designed for a dedicated `ai-team` profile; normal `desktop`/`web` sessions keep their preset roster unless explicitly installed. This machine also has it in `desktop` for UI preview. Multi-Project is concurrent: every Project owns its ledger and runtime, each has its own Waker/SSE/tool routing, and `activeProjectId` only selects the UI default. Full install / uninstall / upgrade steps, data locations and profile-isolation checks: [docs/10-安装卸载升级.md](docs/10-安装卸载升级.md).
+
+## Layout
+
+- `index.js` — Host entry: same-origin `/ai-team/api/health`, `/ai-team/api/bootstrap`, and `POST /ai-team/api/projects/:id/commands` backed by the default-project ledger.
+- `client.js` — Client entry: `sidebar.panellist` icon + `main` key `ai-team` three-pane workbench.
+- `cordis.patch.yml` — bundle row.
+- `lib/` — append-only ledger (`ledger/*`), task/thread/inbox domain (`domain/*`) and the host `TeamService` (`host/service.js`).
+- `test/` — `node --test`: bundle smoke, ledger/fold/checkpoint, domain rules, and a real-ledger service integration.
+
+The bundled SkillOpt integration is derived from Microsoft SkillOpt (`plugins/dsh`, MIT); its license is kept at `lib/vendor/skillopt/LICENSE`.
+
+## Rules
+
+- Only `--dsw-alias-*` tokens for UI chrome; the Loop Mark is the only artwork with its own color.
+- No timer / no background wake; M1 will add event-driven wake.
+- A′ (native session + tool view) is only the fallback when the workbench is unavailable.
+
+## Docs
+
+The phase-1 research, PRD, architecture, UX and execution plan live in the repository's `docs/` directory (index: `docs/README.md`).