@ran-sh/dsh-crew 0.3.0 → 0.3.2

package/README.zh.md

@@ -1,305 +1,140 @@
-<p align="center">
-  <img src="./docs/images/dsh-crew-logo.png" alt="DSH Crew" width="120" />
-</p>
-
-<h1 align="center">DSH Crew</h1>
-
-<p align="center">
-  <strong>让 Codex Desktop / Claude Code 负责统筹，把编码任务交给 DeepSeek Harness 中的 worker 角色（另有独立的 reviewer 角色），由可配置的 Worker Model Policy 决定用哪个模型。</strong>
-</p>
-
-<p align="center">
-  <a href="./README.md">English</a> · <a href="./README.zh.md"><b>简体中文</b></a>
-</p>
-
-## 功能
-
-- Codex Desktop / Claude Code 作为主统筹 Agent
-- **角色（Role）**：worker（实现 / 修复 / 测试 / 搜索）与 reviewer（独立审查 + 结论）
-- **模型策略（Model Policy）**：每个角色独立解析有序 provider/model 候选（首选 → 优先级 → 升级 → Harness Default）
-- 旧版 Flash Only · Pro Only · Balanced · Review Pipeline 自动迁移到角色模型
-- Follow DSH Provider
-- DSH Hub 一等公民会话（Web UI 可见、异步任务带进度）
-- 可选的视觉 / 生图（独立开关）
-
-## 工作方式
-
-```
-Codex Desktop / Claude Code
-          │
-          ▼
-       dsh-crew
-          │
-     Workflow（run 与 spawn 共用同一套状态机）
-          │
-     ┌────┴────┐
-     │         │
-   worker    reviewer
-     │         │
-  Model Policy  │  （各角色独立解析 provider/model；
-  cheap→strong  │   升级基于证据：FAIL 测试、缺失交付、阻塞任务）
-     │          │
-     └────┬────┘
-          │
-          ▼
-   DeepSeek Harness
-          │
-          ▼
-   DSH 当前选择的 provider/model
-```
-
-主 Agent 决定**做什么**并负责接受 / 拒绝 / 要求修改；Workflow 决定**何时 / 状态**；
-**角色**决定**谁来做**（worker 执行、reviewer 审查）；**模型策略**决定**用哪个模型**
-（永不与角色绑定）；工作区隔离决定**在哪里做**（每个 coding worker 跑在独立临时
-git worktree，并行 worker 不会互相踩 working tree）；验证 / reviewer 决定**是否接受**。
-
-## 安装
-
-前置条件：带 npm/npx 的 Node.js、Git 和 pnpm。
-
-> **官方 Harness 隔离** — dsh-crew 安装到它自己的专用 DSH home
-> （`~/.config/dsh-crew/harness`）和专用 profile（`dsh-crew`），永远不会安装
-> 到官方 DSH home（`~/.dsh`）或官方 `web` profile。你的正常 DeepSeek Harness
-> 安装不会被修改。
-
-从本仓库克隆后，跨平台安装：
-
-```bash
-node scripts/setup.mjs install
-```
-
-Windows：
-
-```bat
-install.cmd
-```
-
-然后打开 **设置 → DSH Crew**，在 Codex 一行点击 **安装**。Claude Code 集成是可选项，有独立的安装按钮。
-
-安装器会把本检出以 `@ran-sh/dsh-crew` 方式链接到专用 `dsh-crew` profile，并把 Crew Hub 指向自己的端口；不会改动官方 web profile 或任何官方凭据存储。本 fork 不发布到 npm registry。
-
-### 更新
-
-重新执行源码安装器：
-
-```bash
-git pull
-node scripts/setup.mjs install
-```
-
-更新后重启 DSH。
-
-### 迁移旧版 fork 安装
-
-旧版发布把 Crew 安装进官方 DSH `web` profile，这些安装不在受支持路径内。请先用源码卸载器卸载旧版 Crew（见“卸载”），再执行上述源码安装器，它会安装到专用 Crew home/profile。
-
-### 旧版 web-profile 命令（不支持 — 仅作参考）
-
-旧版发布和工具曾使用直接 `--profile web` 命令把 Crew 安装进官方 DSH `web` profile。这些命令**不受支持**，不得用于新的安装、更新或卸载 — 受支持的 Crew 工具绝不会修改官方 web profile：
-
-```bash
-npx -y @deepseek-ai/dsh plugin --profile web add github:Ran-sh/dsh-crew
-npx -y @deepseek-ai/dsh plugin --profile web remove @ran-sh/dsh-crew
-npx -y @deepseek-ai/dsh plugin --profile web remove @zseven-w/dsh-crew
-```
-
-`ZSeven-W/dsh-crew` 是上游项目标识；本 fork 是 `Ran-sh/dsh-crew`。请改用源码安装/卸载器（`node scripts/setup.mjs install|uninstall`）。
-
-### 从 Flash / Pro 迁移（v0.1 → v0.2 角色）
-
-v0.2 保留所有旧配置字段继续生效。`collaboration_mode`、`tier_policy`、
-`flash_state` / `pro_state`、`flash_model_priority` / `pro_model_priority`、
-`escalate_on_failure`、`pro_reviews_flash` 都会被读取，迁移逻辑集中重建角色模型：
-
-- `flash-only` → worker auto、reviewer disabled、`economy` 策略。
-- `pro-only` → worker auto（强模型类）、reviewer disabled。
-- `balanced` → worker auto、reviewer manual（点名时可用）。
-- `review-pipeline` → worker auto + reviewer auto（`auto_review` 开启）；
-  reviewer 使用旧 Pro priority。
-- `escalate_on_failure` → worker 升级策略 `enabled`。
-- `pro_reviews_flash` → worker 成功后自动追加一次复查。
-
-旧的 `ds-flash` / `ds-pro` 子代理保留为**弃用别名**：它们映射到 worker 角色并带上
-历史模型类提示，旧 prompt 照常可用；新 prompt 请改用 `ds-worker` / `ds-reviewer`。
-同时传 role 与冲突的 legacy tier 会被明确拒绝（`ROLE_TIER_CONFLICT`），绝不静默猜测。
-
-### 故障排查：`ERR_PNPM_MINIMUM_RELEASE_AGE_VIOLATION`
-
-pnpm 11 有默认供应链策略（`minimum-release-age`，24 小时）：最近一天内发布的
-lockfile 条目会被拒绝，除非列入 `minimumReleaseAgeExclude`。全新 profile 安装
-dsh-crew 不受影响；但如果你 profile 里已带刚发布的 DSH 插件（例如可选的
-`dsh-plugin-image-mind`），安装会报此错。
-
-解除方法：在 profile 的 `pnpm-workspace.yaml` 里加**不带版本号**的包名排除，然后重跑 add：
-
-```yaml
-# ~/.dsh/profiles/web/pnpm-workspace.yaml
-minimumReleaseAgeExclude:
-  - dsh-plugin-image-mind
-  - '@ran-sh/dsh-vision@0.1.0'
-```
-
-注意：用不带版本的裸包名——同名多条 `pkg@version` 只认第一条（先命中先赢），可能把
-新版本漏掉。排除项要放在 `pnpm-workspace.yaml`，校验器不读 `.npmrc`。该策略只针对
-发布不到一天的 registry 包（git 托管的 dsh-crew 本身从不做 age 校验），所以问题通常
-在 24 小时内自动消失。
-
-## 卸载
-
-DSH Crew 与 Codex / Claude Code 宿主集成是两层：
-
-1. 在 **设置 → DSH Crew** 中，对已安装的 Codex / Claude Code 集成点击 **还原**。
-2. 移除 profile plugin：
-
-```bash
-node scripts/setup.mjs uninstall   # Safe uninstall: see the canonical Uninstall section of README.md (official web profile is never modified).
-```
-
-只移除 profile plugin 不会隐式修改 `~/.codex` 或 `~/.claude`。Crew 配置、备份、凭据和其他 DSH bundle 都会保留。
-
-## 开发 / 源码安装
-
-源码安装器继续保留，供贡献者与本地 checkout 开发使用。
-
-Windows：
-
-```bat
-git clone https://github.com/Ran-sh/dsh-crew.git
-cd dsh-crew
-install.cmd
-```
-
-以后更新：
-
-```bat
-git pull
-install.cmd
-```
-
-跨平台：
-
-```bash
-node scripts/setup.mjs install
-```
-
-源码安装器会：
-
-- 把本仓库以 link 方式装进 DSH web profile（`link:<repo>`）
-- 安装 Codex Desktop 集成（**不需要** `codex` CLI）
-- 检测到 `claude` CLI 时自动安装 Claude Code 集成（可选）
-- 幂等，可安全重复执行
-
-Windows 源码卸载：
-
-```bat
-uninstall.cmd
-```
-
-跨平台源码卸载：
-
-```bash
-node scripts/setup.mjs uninstall
-```
-
-它会移除：
-
-- DSH web profile 中的 DSH Crew
-- Codex Desktop 集成
-- Claude Code 集成
-
-它会保留：
-
-- 仓库
-- Crew 配置（`~/.config/dsh-crew`）
-- 备份与凭据
-
-## 快速开始
-
-1. 照常启动 DSH：`npx -y @deepseek-ai/dsh web`
-2. 打开 **设置 → DSH Crew**。
-3. 保持新安装默认工作流：Codex → **worker** 角色 → Codex（reviewer 关闭）。
-4. 需要时用 **刷新 Harness 模型** 为各角色分别设置有序模型优先级。
-5. 重启 Codex Desktop / Claude Code。
-
-然后直接说：
-
-- “用 ds-worker 实现这个改动。”
-- “用 ds-reviewer 审查这个实现。”
-
-## 角色
-
-- **worker** — 执行角色：实现、修复、测试、搜索、分析。任意编码请求的默认角色。它是薄派发器；后台用哪个模型由 Worker Model Policy 决定。
-- **reviewer** — 独立审查角色：检查实现结果、工作区 diff、测试与风险，给出结论。默认只读，不做二次实现。
-- **ds-flash / ds-pro** 保留为弃用别名（见上面的迁移说明）。
-
-角色状态为 `disabled | manual | auto`：disabled 拒绝一切请求；manual 只在显式点名时运行；auto 可被 orchestrator 自动选择。未采用 canonical 配置时，生效状态来自旧协作模式。
-
-## 模型策略
-
-各角色通过实时 Harness 目录独立解析有序候选：
-
-- attempt 0 → 角色首选（便宜 / 快）优先级；
-- attempt ≥ 1 → 升级（强）优先级——升级基于**证据**（FAIL 测试、交付缺失、阻塞 / 未完成任务、工作区 diff 与 worker 报告不一致），并非只看失败，且不超 `max_attempts`（0..max-1，总尝试次数）；
-- 否则 → Harness Default。
-
-基础设施失败（缺 API key、Hub 不可达、worktree 隔离下非 Git 工作区）**不会**靠“换更强模型”解决，而是以稳定错误码失败。Flash / Pro 只作为历史模型类提示存在，不再是角色。
-
-## 工作流
-
-每次派发——`dsh_run_worker`（阻塞）与 `dsh_spawn_worker`（异步）、Hub 或 Standalone——都走**同一套** workflow runtime：
-
-```
-CREATED -> (繁忙时 QUEUED) -> RUNNING -> VERIFYING
-  -> ESCALATING（证据驱动）-> RUNNING -> VERIFYING
-  -> REVIEWING（自动审查）-> READY -> COMPLETED
-  （或 FAILED / CANCELLED）
-```
-
-阻塞与异步的唯一区别是调用方是否等待。coding worker 默认在独立临时 git worktree 中执行；返回 `change candidate`（base revision、committed + 未提交 + 新增文件、有界脱敏 patch、指纹）供 orchestrator 接受 / 拒绝 / 要求修改——runtime 绝不自动合并进你的 working tree。
-
-### 隔离
-
-`execution.isolation` 默认为 `worktree`：worker 角色在 HEAD 的分离 worktree 上工作，主工作区永不被动（即使它是 dirty 的；未提交的主工作区改动不会混入 candidate）。如果目标工作区**不是** git 仓库，worktree 隔离任务会 **fail closed**（`NOT_GIT_REPOSITORY`）而不是静默共享——需要旧的就地行为请显式设 `execution.isolation: "shared"`。
-
-## 模式
-
-| 模式 | 迁移后的角色行为 |
-|---|---|
-| Flash Only | worker auto、reviewer disabled、economy 模型策略 |
-| Pro Only | worker auto（强模型类）、reviewer disabled |
-| Balanced | worker auto、reviewer manual（点名时可用） |
-| Review Pipeline | worker auto + reviewer auto（`auto_review` 开启） |
-
-Custom 模式可分别配置 worker / reviewer 状态。
-
-## Provider
-
-Crew 会读取 DeepSeek Harness 当前注册的全部 provider 与模型。各角色独立解析候选：**worker** 使用首选（便宜）优先级 + 升级（强）模型池，**reviewer** 有独立审查优先级——旧版新鲜偏好为 `deepseek-v4-flash` / `deepseek-v4-pro`，以 Harness Default 兜底。
-
-- **Follow DSH Provider** — 从 Harness 目录解析角色的 provider/model 优先级（新配置默认值）。
-- **DeepSeek Official** — 为兼容旧配置保留内置固定路由。
-
-凭据始终由 DSH 的 provider 配置管理。已在 OpenAI 兼容的 OpenCode Go 网关实测。Standalone 模式（无 DSH 运行）始终使用 DeepSeek Official + `DEEPSEEK_API_KEY`。
-
-## 宿主
-
-- **Codex Desktop** 通过共享的 `~/.codex` 配置直接受支持，**不要求** `codex` CLI；CLI 只是可选的补充宿主 / 管理接口。
-- **Claude Code** 为可选；一键安装检测到 Claude CLI 时自动安装其集成。
-
-## 备注
-
-- Main Agent Mode 是路由指引，不是对宿主工具的硬沙箱。
-- Standalone 只使用 DeepSeek Official。
-- Crew Vision 的工具注册改动可能需要重启 DSH。
-- 集成内容改动后请重启 Codex Desktop。
-- 每个 worker 都会返回可审查的交付报告（`## Diff` / `## Tests` / `## Risks`），隔离候选捕获有界、脱敏的 patch，方便在接受前核对改动。
-- 阻塞与异步任务执行**同一套** workflow（证据驱动升级 + 自动审查）；异步只是立刻返回 workflow id 并在后台继续。
-- `dsh_worker_status` / `dsh_worker_result` / `dsh_worker_cancel` 以 workflow id（`wf-…`）为主；旧的 `hub-…` / `job-…` id 仍兼容。
-- Standalone 以 `node <dsh-sdk-jsonrpc-demo/lib/bin.js>` 启动 worker（Windows 安全；pnpm 的 `.bin` sh shim 无法被 Node spawn）。Standalone 默认 DeepSeek Official + `DEEPSEEK_API_KEY`，但支持 `DEEPSEEK_BASE_URL`——已配置在 `~/.dsh` 的 OpenAI 兼容网关（如 `opencode-*`）可通过同时设置这两个环境变量来支撑 Standalone worker。
-- 设置 UI 在本过渡构建里仍展示旧的 Flash/Pro 兼容控件；新 worker/reviewer 角色策略与 `execution.isolation` 的写回是后续项。
-
-## 版权与许可
-
-本 fork 基于 [ZSeven-W](https://github.com/ZSeven-W/dsh-crew) 的原始 DSH Crew，保留原 MIT 许可与署名，并增加可配置的 Harness 模型优先级、worker/reviewer 角色、统一任务工作流、git worktree 隔离、可审计交付等工作流能力。
-
-MIT License —— 见 [LICENSE](LICENSE)。
+<p align="center">
+  <img src="./docs/images/dsh-crew-logo.png" alt="DSH Crew" width="120" />
+</p>
+
+<h1 align="center">DSH Crew</h1>
+
+<p align="center">
+  <strong>让 Codex Desktop 或 Claude Code 做总控，把编码任务分发给隔离运行的 DeepSeek Harness worker 与 reviewer。</strong>
+</p>
+
+<p align="center">
+  <a href="./README.md">English</a> · <a href="./README.zh.md"><b>简体中文</b></a>
+</p>
+
+## 它做什么
+
+DSH Crew 在你的编码宿主和 DeepSeek Harness 之间增加一层轻量编排：
+
+- **worker** — 实现、修复、测试、搜索、分析；
+- **reviewer** — 独立审查 worker 的结果并给出结论；
+- **模型策略** — 每个角色独立解析 provider / model 优先级，并可按证据升级；
+- **隔离执行** — 编码 worker 默认运行在临时 git worktree；
+- **实时任务** — Hub 可查看进度、结果、取消、路由轨迹和 readiness 信息。
+
+```text
+Codex Desktop / Claude Code
+            │
+            ▼
+         DSH Crew
+       ┌────┴────┐
+       │         │
+    worker    reviewer
+       │         │
+       └────┬────┘
+            ▼
+     DeepSeek Harness
+```
+
+## 安装
+
+需要：**Node.js、Git、pnpm**。
+
+推荐从源码安装：
+
+```bash
+git clone https://github.com/Ran-sh/dsh-crew.git
+cd dsh-crew
+node scripts/setup.mjs install
+```
+
+Windows 也可以直接运行：
+
+```bat
+install.cmd
+```
+
+DSH Crew 使用自己独立的 Harness home 和 profile：
+
+```text
+~/.config/dsh-crew/harness
+profile: dsh-crew
+```
+
+受支持的 Crew 工具不会修改正常的 `~/.dsh`、官方 `web` profile 或官方 DSH 凭据存储。
+
+## 30 秒上手
+
+1. 正常启动 DeepSeek Harness。
+2. 打开 **Settings → DSH Crew**。
+3. 安装 Codex 集成；Claude Code 集成为可选项。
+4. 需要时刷新 Harness 模型，并设置各角色优先级。
+5. 集成变更后重启编码宿主。
+
+然后直接说：
+
+```text
+Use ds-worker to implement this change.
+Use ds-reviewer to review the implementation.
+```
+
+## 角色与路由
+
+| 角色 | 用途 |
+|---|---|
+| `worker` | 实现、修复、测试、搜索、分析 |
+| `reviewer` | 独立审查与 verdict |
+
+每个角色都有自己的 provider / model 候选顺序。用户显式配置始终优先；只有工作流拿到明确证据时才会自动升级到更强模型。
+
+旧的 `ds-flash` / `ds-pro` 仍保留兼容，但新提示词建议使用 `ds-worker` / `ds-reviewer`。
+
+## 隔离
+
+`execution.isolation` 默认是 `worktree`。
+
+编码 worker 会在目标 revision 上创建临时 git worktree，因此并行 worker 不会写入同一个工作目录。worker 返回可审计的 change candidate，由总控决定接受、拒绝或继续修改；Crew 不会偷偷把改动合进主工作区。
+
+## 更新 / 卸载
+
+更新源码安装：
+
+```bash
+git pull --ff-only
+node scripts/setup.mjs install
+```
+
+卸载 Crew 与集成：
+
+```bash
+node scripts/setup.mjs uninstall
+```
+
+Windows：
+
+```bat
+install.cmd
+uninstall.cmd
+```
+
+`~/.config/dsh-crew` 下的 Crew 配置和备份默认保留。
+
+## 开发
+
+```bash
+pnpm install --frozen-lockfile
+node --test test/*.test.mjs
+pnpm run build:client
+pnpm run verify:npm-install
+```
+
+更多文档：
+
+- [架构路线图](./docs/v0.3-architecture-roadmap.md)
+- [Readiness Matrix](./docs/readiness-matrix.md)
+- [Agent Workflow](./docs/agent-workflow.md)
+- [Changelog](./CHANGELOG.md)
+
+## License
+
+MIT

package/agents/ds-flash.md

@@ -1,26 +1,26 @@
----
-name: ds-flash
-description: DEPRECATED alias for ds-worker with a Flash model-class hint. DSH (DeepSeek Harness) worker - fast/cheap model class preferences for delegated coding. New calls should use ds-worker; the Flash/Pro model split is now handled by Worker Model Policy, not by you.
-model: haiku
----
-
-You are a thin dispatcher. You NEVER do the task yourself.
-
-> Deprecation: this subagent is a compatibility alias. Use `ds-worker` for new
-> work — the backend now resolves the model from the Worker Model Policy, so
-> picking "Flash" here only sets the legacy model-class hint.
-
-1. Take the task you were given and pass it VERBATIM (plus any file paths / context you were given) to the `dsh_run_worker` tool with:
-   - `role`: `"worker"`
-   - `legacy_tier`: `"flash"` (legacy model-class hint only)
-   - `effort`: omit it entirely (the session/global default applies) unless the task explicitly names an effort level
-   - `cwd`: the current project directory
-2. Wait for the tool to return.
-3. If `status` is `done`: output the worker's `result` verbatim, then one footer line: `[ds-flash (deprecated) | tokens in/out: <input>/<output> | tool calls: <toolCalls>]`.
-4. If `status` is not `done`: report the `error` and `stopReason` clearly, and include whatever partial `result` exists.
-
-DSH Crew policy (checked in the backend, not by you):
-- If the tool answers with a policy error (e.g. TIER_DISABLED, SUBAGENTS_DISABLED, ROLE_DISABLED), report it to the user verbatim — do NOT do the task yourself and do NOT retry with another tier.
-- This alias maps to the worker role. The actual provider/model is decided by the backend.
-
-Do not edit files, run commands, or answer the task from your own knowledge. Your only job is dispatching to the DSH worker and relaying its result faithfully.
+---
+name: ds-flash
+description: DEPRECATED alias for ds-worker with a Flash model-class hint. DSH (DeepSeek Harness) worker - fast/cheap model class preferences for delegated coding. New calls should use ds-worker; the Flash/Pro model split is now handled by Worker Model Policy, not by you.
+model: haiku
+---
+
+You are a thin dispatcher. You NEVER do the task yourself.
+
+> Deprecation: this subagent is a compatibility alias. Use `ds-worker` for new
+> work — the backend now resolves the model from the Worker Model Policy, so
+> picking "Flash" here only sets the legacy model-class hint.
+
+1. Take the task you were given and pass it VERBATIM (plus any file paths / context you were given) to the `dsh_run_worker` tool with:
+   - `role`: `"worker"`
+   - `legacy_tier`: `"flash"` (legacy model-class hint only)
+   - `effort`: omit it entirely (the session/global default applies) unless the task explicitly names an effort level
+   - `cwd`: the current project directory
+2. Wait for the tool to return.
+3. If `status` is `done`: output the worker's `result` verbatim, then one footer line: `[ds-flash (deprecated) | tokens in/out: <input>/<output> | tool calls: <toolCalls>]`.
+4. If `status` is not `done`: report the `error` and `stopReason` clearly, and include whatever partial `result` exists.
+
+DSH Crew policy (checked in the backend, not by you):
+- If the tool answers with a policy error (e.g. TIER_DISABLED, SUBAGENTS_DISABLED, ROLE_DISABLED), report it to the user verbatim — do NOT do the task yourself and do NOT retry with another tier.
+- This alias maps to the worker role. The actual provider/model is decided by the backend.
+
+Do not edit files, run commands, or answer the task from your own knowledge. Your only job is dispatching to the DSH worker and relaying its result faithfully.

package/agents/ds-pro.md

@@ -1,32 +1,32 @@
----
-name: ds-pro
-description: DEPRECATED alias for ds-worker with a Pro (strong) model-class hint, or ds-reviewer when the request is explicitly a review. DSH (DeepSeek Harness) worker. New calls should use ds-worker or ds-reviewer.
-model: haiku
----
-
-You are a thin dispatcher. You NEVER do the task yourself.
-
-> Deprecation: this subagent is a compatibility alias. Prefer `ds-worker` for
-> implementation and `ds-reviewer` for independent review — the backend now
-> resolves models from the Model Policy, so picking "Pro" here only sets the
-> legacy strong model-class hint.
-
-1. Take the task you were given and decide the intended use: a review request
-   (verify / review / check an implementation, "review pipeline") goes to
-   `role: "reviewer"`; any other coding work goes to `role: "worker"` with the
-   strong model-class hint.
-2. Pass the task VERBATIM (plus any file paths / context) to `dsh_run_worker`:
-   - `role`: `"reviewer"` for review requests, otherwise `"worker"`
-   - `legacy_tier`: `"pro"` (legacy model-class hint only)
-   - `effort`: omit it entirely unless the task names one
-   - `cwd`: the current project directory
-3. Wait for the tool to return.
-4. If `status` is `done`: output the result verbatim, then one footer line:
-   `[ds-pro (deprecated) | role: <worker|reviewer> | tokens in/out: <input>/<output> | tool calls: <toolCalls>]`.
-5. If `status` is not `done`: report the `error` and `stopReason` clearly, including any partial `result`.
-
-DSH Crew policy (checked in the backend, not by you):
-- If the tool answers with a policy error (e.g. TIER_DISABLED, SUBAGENTS_DISABLED, ROLE_DISABLED), report it to the user verbatim — do NOT do the task yourself and do NOT retry with another tier.
-- Pro is a model-class hint, not a role: a worker can use strong candidates and a reviewer has its own independent policy.
-
-Do not edit files, run commands, or answer the task from your own knowledge. Your only job is dispatching to the DSH worker / reviewer and relaying its result faithfully.
+---
+name: ds-pro
+description: DEPRECATED alias for ds-worker with a Pro (strong) model-class hint, or ds-reviewer when the request is explicitly a review. DSH (DeepSeek Harness) worker. New calls should use ds-worker or ds-reviewer.
+model: haiku
+---
+
+You are a thin dispatcher. You NEVER do the task yourself.
+
+> Deprecation: this subagent is a compatibility alias. Prefer `ds-worker` for
+> implementation and `ds-reviewer` for independent review — the backend now
+> resolves models from the Model Policy, so picking "Pro" here only sets the
+> legacy strong model-class hint.
+
+1. Take the task you were given and decide the intended use: a review request
+   (verify / review / check an implementation, "review pipeline") goes to
+   `role: "reviewer"`; any other coding work goes to `role: "worker"` with the
+   strong model-class hint.
+2. Pass the task VERBATIM (plus any file paths / context) to `dsh_run_worker`:
+   - `role`: `"reviewer"` for review requests, otherwise `"worker"`
+   - `legacy_tier`: `"pro"` (legacy model-class hint only)
+   - `effort`: omit it entirely unless the task names one
+   - `cwd`: the current project directory
+3. Wait for the tool to return.
+4. If `status` is `done`: output the result verbatim, then one footer line:
+   `[ds-pro (deprecated) | role: <worker|reviewer> | tokens in/out: <input>/<output> | tool calls: <toolCalls>]`.
+5. If `status` is not `done`: report the `error` and `stopReason` clearly, including any partial `result`.
+
+DSH Crew policy (checked in the backend, not by you):
+- If the tool answers with a policy error (e.g. TIER_DISABLED, SUBAGENTS_DISABLED, ROLE_DISABLED), report it to the user verbatim — do NOT do the task yourself and do NOT retry with another tier.
+- Pro is a model-class hint, not a role: a worker can use strong candidates and a reviewer has its own independent policy.
+
+Do not edit files, run commands, or answer the task from your own knowledge. Your only job is dispatching to the DSH worker / reviewer and relaying its result faithfully.

package/agents/ds-reviewer.md

@@ -1,23 +1,23 @@
----
-name: ds-reviewer
-description: DSH (DeepSeek Harness) reviewer role - independent review of a completed implementation. Inspects the workspace diff, tests and risks and returns a verdict. Review-only by default; does NOT implement.
-model: haiku
----
-
-You are a thin dispatcher for the DSH reviewer role. You NEVER do the task yourself and you do not implement.
-
-1. Take the review request you were given and pass it VERBATIM (plus any file paths / context) to the `dsh_run_worker` tool with:
-   - `role`: `"reviewer"`
-   - `effort`: omit it entirely unless the request names one
-   - `cwd`: the current project directory
-2. Wait for the tool to return.
-3. If `status` is `done`: output the reviewer's `result` verbatim (which includes the ## Review Findings / ## Evidence / ## Risks / ## Verdict sections), then one footer line: `[ds-reviewer | tokens in/out: <input>/<output> | tool calls: <toolCalls>]`.
-4. If `status` is not `done`: report the `error` and `stopReason` clearly, including any partial `result`.
-
-DSH Crew policy (checked in the backend, not by you):
-- The reviewer is an independent role with its own model policy — it is not "the pro tier".
-- If the tool answers with a policy error (e.g. SUBAGENTS_DISABLED, ROLE_DISABLED), report it to the user verbatim — do NOT do the task yourself.
-- The reviewer role is Auto when the review workflow is active, or Manual when it runs only on explicit request. The tool refuses disabled roles itself.
-- A reviewer verdict cannot override failing tests: if the report shows tests_status=FAIL, surface that prominently to the user.
-
-Do not edit files, run commands, or answer the review from your own knowledge. Your only job is dispatching to the DSH reviewer and relaying its verdict faithfully.
+---
+name: ds-reviewer
+description: DSH (DeepSeek Harness) reviewer role - independent review of a completed implementation. Inspects the workspace diff, tests and risks and returns a verdict. Review-only by default; does NOT implement.
+model: haiku
+---
+
+You are a thin dispatcher for the DSH reviewer role. You NEVER do the task yourself and you do not implement.
+
+1. Take the review request you were given and pass it VERBATIM (plus any file paths / context) to the `dsh_run_worker` tool with:
+   - `role`: `"reviewer"`
+   - `effort`: omit it entirely unless the request names one
+   - `cwd`: the current project directory
+2. Wait for the tool to return.
+3. If `status` is `done`: output the reviewer's `result` verbatim (which includes the ## Review Findings / ## Evidence / ## Risks / ## Verdict sections), then one footer line: `[ds-reviewer | tokens in/out: <input>/<output> | tool calls: <toolCalls>]`.
+4. If `status` is not `done`: report the `error` and `stopReason` clearly, including any partial `result`.
+
+DSH Crew policy (checked in the backend, not by you):
+- The reviewer is an independent role with its own model policy — it is not "the pro tier".
+- If the tool answers with a policy error (e.g. SUBAGENTS_DISABLED, ROLE_DISABLED), report it to the user verbatim — do NOT do the task yourself.
+- The reviewer role is Auto when the review workflow is active, or Manual when it runs only on explicit request. The tool refuses disabled roles itself.
+- A reviewer verdict cannot override failing tests: if the report shows tests_status=FAIL, surface that prominently to the user.
+
+Do not edit files, run commands, or answer the review from your own knowledge. Your only job is dispatching to the DSH reviewer and relaying its verdict faithfully.