dsh-grill-first 0.1.3 → 0.1.5

package/CHANGELOG.md

@@ -2,6 +2,43 @@
 
 本项目遵循 [语义化版本](https://semver.org/lang/zh-CN/)。
 
+## 0.1.5
+
+"让人一眼看懂"版本：只有文档，没有代码改动。
+
+### 新增
+
+- **英文 README**（`README.en.md`）：面向英文用户重写的，不是逐字翻译——一句话卖点、30 秒对话演示、四档对比、配置表、兼容性与已知限制。
+- **中文 README 首屏重做**：一句话主张「Agent 不是不会做，是经常做完了你才发现不是你要的」、一段真实对话示例、四种门档位的对比表，都提到首屏。
+
+### 变更
+
+- 中英文 README 顶部互相加了语言切换链接。
+
+## 0.1.4
+
+"真的按你确认过的意思做"版本：修掉四个"配置写了但不生效"的地方，并让询问类消息不再被当成模糊任务。
+
+### 新增
+
+- **`gateMode: "ask"` 档位**：需求没确认时，变更类操作弹一张 DSH 原生确认卡片问你"照现在的理解直接动手吗"，你点头就放行——不想切 `strict` 硬门禁、又不想写配置文件的人用这个。卡片只弹一次，之后本次任务照常跑。
+- **`memoryScope` 配置**：项目背景与历史默认存工作区；改成 `global` 则落到全局目录的 `project.json`，跨项目共享。
+- **问句不再触发澄清**：问号、"怎么样 / 查一下 / 看看 / 解释一下 / 我想知道" 这类询问，以及 `look at / check my / tell me / explain` 这类英文问句，一律直接放行并接话作答——0.1.3 及以前会被当成"模糊任务"要求先澄清。
+- **`intensity` 真正生效**：`gentle` 两维齐就放行、`standard` 三维、`grill` 还要求写清验收标准（以前三个档位阈值完全一样）。
+- **英文任务也能判**：四张维度词表补了英文同义词（write/implement/refactor、only/scope、use/based on、test/verify…），英文按词边界匹配，不会把 `skipList` 误当跳过词。
+
+### 修复
+
+- **`maxRounds` 现在真的会到上限**：以前 `grill_start` 每次都把轮数清零，问多少轮都不会停；现在问满上限后直接放行，并在返回值里给出 `rounds_used`。
+- **用户追问不再把门禁关回去**：以前任何一条用户消息都会重置门禁，"继续""嗯""好的"也会；现在只把**新任务**当成新一轮，追问/答话不会打断进行中的工作。
+- **历史记录不再无限增长**：项目历史只保留最近 100 条（摘要只用最近 3 条，多的读不到只是白占地方）。
+- **删掉 `grill_memory` 这个不存在的只读白名单条目**（0.1.3 之前遗留的死引用）。
+
+### 内部
+
+- 单测/端到端用例从 94 个补到 127 个，新增：问句识别、intensity 阈值、`ask` 档位、追问不重置门禁、`maxRounds` 上限、`memoryScope`、历史裁剪，以及一张"主人真实说过的 11 句话"回归样例表。
+- CI 增加 `npm run typecheck`；新增 `.github/workflows/release.yml`：打 `v*` tag 自动发 npm（需要仓库 Secrets 配 `NPM_TOKEN`）。
+
 ## 0.1.3
 
 文档修正版。

package/LICENSE

@@ -1,21 +1,21 @@
-MIT License
-
-Copyright (c) 2026 InoriMutsumi
-
-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.
+MIT License
+
+Copyright (c) 2026 InoriMutsumi
+
+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.en.md

@@ -0,0 +1,167 @@
+# dsh-grill-first
+
+![tests](https://img.shields.io/badge/tests-127%20passing-brightgreen)
+![license](https://img.shields.io/badge/license-MIT-blue)
+![DSH plugin](https://img.shields.io/badge/DSH-plugin-blueviolet)
+
+> A [DSH (DeepSeek Harness)](https://github.com/deepseek-ai/dsh) plugin that **interviews you before it starts working**: vague requests get clarified first, and the main task only starts once you confirm the summary.
+
+**Agents rarely fail because they can't do the work. They fail because they do the wrong work, confidently.**
+
+[中文](./README.md) · [Install](#install) · [Configuration](#configuration) · [Known limitations](#known-limitations)
+
+## See it in 30 seconds
+
+```text
+you  optimize that page
+
+it   1) Which page? 2) "Optimize" meaning speed, style or copy?
+     3) How do we know it's done?
+
+you  the login page, mostly styling — the alignment and spacing look off
+
+it   GOAL      tidy up the login page's visuals
+     SCOPE     styling only; structure and copy stay as-is
+     CONSTRAINTS use the existing design tokens
+     ACCEPTANCE key screens don't break on mobile
+                          [ Confirm and start ]  [ Change something ]
+
+you  confirm
+
+it   (only now does it touch anything)
+```
+
+Without the plugin, the same sentence usually goes like this:
+
+```text
+you  optimize that page
+
+it   Done! I've updated the homepage copy and color scheme ✨   ← you meant the login page
+```
+
+## The four gate modes
+
+| | No plugin | `gateMode: warn` (default) | `gateMode: ask` | `gateMode: strict` |
+|---|---|---|---|---|
+| Vague request | Guesses and starts | Asks; actions not blocked | Asks; one confirmation card before acting | Asks; blocks writes and commands until confirmed |
+
+`warn` is the default so upgrading never silently freezes a workflow you already had.
+
+## How it works
+
+Every new task starts with a check, not with work:
+
+1. The model passes your request to `grill_start`.
+2. If it's missing goal / scope / constraints / acceptance criteria, the plugin asks up to 3 questions per round, 3 rounds max, each round built on your previous answer.
+3. Once the picture is clear, `grill_confirm` hands you a four-part summary (goal / scope / constraints / acceptance criteria).
+4. "Confirm" releases the task; "change something" makes it revise and ask again.
+5. One-liners and questions don't trigger any of this — nor do messages containing `直接做`, `跳过提问`, `不用问`, `skip`, `skip the questions`, `just do it`.
+
+## Install
+
+Quit any running task first. Any of these works:
+
+**Plugin marketplace (recommended)** — DSH settings → Plugins → search `dsh-grill-first`.
+
+**npm**
+
+```bash
+dsh plugin --profile <your-profile> add dsh-grill-first
+```
+
+**From source**
+
+```bash
+dsh plugin --profile <your-profile> add github:InoriMutsumi/dsh-grill-first
+```
+
+Live profiles hot-load the plugin — no restart needed (restart DSH once if it doesn't show up).
+
+## What you get
+
+- **A protocol paragraph** injected into the system prompt (~600 chars) telling the model to run `grill_start` before new work. If you can see `【开工前提问协议（grill-first）】` in the prompt, it's installed.
+- **Two tools:**
+
+| Tool | What it does |
+|---|---|
+| `grill_start` | Decides whether the request needs clarification. Returns `verdict`, `focus_hints` (what to ask), and `memory_context` (preferences you already confirmed, so it doesn't ask twice) |
+| `grill_confirm` | Hands you the four-part summary; on confirmation releases the task and saves what you confirmed |
+
+## When you don't want to be asked
+
+- Put `直接做`, `跳过提问`, `不用问`, `skip`, `skip the questions` or `just do it` in the message → it starts immediately.
+- Asks a question ("how is this plugin", "check my repos", "explain this code") → treated as a question, not a task, so nothing is clarified.
+- Or turn `enabled` off / set `gateMode: off` in the config → fully silent.
+
+## Configuration
+
+DSH settings → Plugins → dsh-grill-first, eight options:
+
+| Option | Type | Default | Meaning |
+|---|---|---|---|
+| `enabled` | boolean | `true` | Master switch. Off = no questions, no gate |
+| `gateMode` | `off` / `warn` / `ask` / `strict` | `warn` | `off` never blocks; `warn` only logs; `ask` shows one confirmation card; `strict` blocks mutating tools until confirmed |
+| `maxQuestionsPerRound` | 1–6 | `3` | Questions per round |
+| `maxRounds` | 1–10 | `3` | Max clarification rounds; on the last round `grill_start` passes the task through |
+| `intensity` | `gentle` / `standard` / `grill` | `standard` | How much detail counts as clear: `gentle` needs two dimensions, `standard` three, `grill` also requires acceptance criteria |
+| `bypassKeywords` | string[] | `["直接做","跳过提问","不用问","skip"]` | Hit any of these → skip the interview |
+| `memoryEnabled` | boolean | `true` | Enable the two-layer memory |
+| `memoryScope` | `project` / `global` | `project` | Where project context lives: in the workspace, or in the global directory shared across projects |
+
+### Editing the config
+
+The plugin page only offers an enable/disable toggle, so edit the profile's patch layer, `<profile>/cordis.patch.yml`:
+
+```yaml
+- id: dsh-grill-first
+  name: dsh-grill-first
+  config:
+    gateMode: ask
+    maxQuestionsPerRound: 3
+    maxRounds: 3
+    intensity: standard
+    memoryEnabled: true
+    memoryScope: project
+    bypassKeywords: [直接做, 跳过提问, 不用问, skip]
+```
+
+Only the keys you set are overridden; the rest keep their defaults. Config changes hot-reload, code changes need a DSH restart.
+
+## Two-layer memory
+
+Only what you actually confirmed is stored — never raw conversation, never the summary as a file.
+
+| Layer | Location | Contents |
+|---|---|---|
+| Global | `<profile>/grill-first/global.json` (falls back to `~/.dsh-grill-first/global.json`) | Standing preferences, e.g. "comments in English" |
+| Project | `<workspace>/.dsh/grill-first/memory.json` | Project background and task history (task name, summary, date; last 100 entries) |
+
+Set `memoryScope: global` to keep project context in the global directory instead — handy when you don't want a `.dsh/` folder in every repo. Next interview, `grill_start` returns this as `memory_context`, so the same question is never asked twice.
+
+## Development
+
+```bash
+npm install                 # dev dependencies (schemastery comes from the host at runtime)
+npm run build               # TypeScript → dist/
+npm test                    # build + 127 unit / end-to-end tests
+node scripts/selfcheck.mjs  # static check: Chinese JSDoc, `any` usage, spec-to-code map
+```
+
+`dist/` is committed on purpose so `dsh plugin add github:…` works without compiling on your machine. CI runs `typecheck` + tests on Node 22.
+
+## Compatibility
+
+- Peer dependency: only `@deepseek-ai/schemastery` (`^3.18.1`, shipped by the host).
+- No import of any `@deepseek-ai/dsh*` runtime package, so no `engines.dsh` and no DSH peer range to keep in sync. Verified against DSH `0.2.1-rc.2`.
+
+## Known limitations
+
+- The gate only applies to the **root agent** that talks to you. Delegated work isn't gated (but delegation tools themselves are, so a subagent can't be used to sidestep it). This is deliberate.
+- `ask` mode asks **once per task** — otherwise a task that writes 20 files would show 20 cards. Use `strict` if you want every action blocked.
+- The judge is **keyword heuristics**, not a semantic model: questions are detected by punctuation and interrogatives. Rare ambiguous sentences can be misjudged; say `直接做` or let the model call `grill_start` again to correct it.
+- The judge runs in Chinese and English only. English keywords match on word boundaries (`skip` matches `skip` and `(skip)` but not `skipList`).
+- Memory is plain local JSON, unencrypted.
+
+## License
+
+MIT — see [LICENSE](./LICENSE).

package/README.md

@@ -1,11 +1,51 @@
 # dsh-grill-first
 
-![tests](https://img.shields.io/badge/tests-91%20passing-brightgreen)
+![tests](https://img.shields.io/badge/tests-127%20passing-brightgreen)
 ![license](https://img.shields.io/badge/license-MIT-blue)
 ![DSH plugin](https://img.shields.io/badge/DSH-plugin-blueviolet)
 
 > DSH（DeepSeek Harness）插件：**开工前先刨根问底**——把模糊需求问成明确需求，你在对话里点头之后，主任务才开工。
 
+**Agent 不是不会做，是经常做完了你才发现不是你要的。**
+
+[English](./README.en.md) · [安装](#安装) · [配置项](#配置项) · [已知限制](#已知限制)
+
+## 30 秒看懂
+
+```text
+你  把那个页面优化一下
+
+它  1) 是哪个页面？2) 「优化」指速度、样式还是文案？
+    3) 怎么才算优化好了？
+
+你  登录页，主要是样式，现在对齐和间距看着乱
+
+它  【目标】整理登录页视觉
+    【范围】只改样式，不动结构和文案
+    【约束】沿用现有设计规范
+    【验收标准】主要页面在移动端不错位
+                     [ 确认，按此开工 ]  [ 要改 ]
+
+你  确认
+
+它  （这时才动手）
+```
+
+不装这个插件时，同一句话大概率是这样：
+
+```text
+你  把那个页面优化一下
+它  好的，已经帮你调整了首页的文案和配色 ✨   ← 你要的是登录页
+```
+
+## 四种档位的区别
+
+| | 不装插件 | `gateMode: warn`（默认） | `gateMode: ask` | `gateMode: strict` |
+|---|---|---|---|---|
+| 任务模糊时 | 直接猜着做 | 追问；动手不拦 | 追问；动手前弹一次确认卡片 | 追问；没确认前拦截写文件、执行命令 |
+
+`warn` 是默认值：升级插件不会突然卡住你已有的工作流。
+
 ## 它解决什么问题
 
 AI 助手最常见的失败不是"不会做"，而是"做完了才发现做的不是你要的"。这个插件在**任何新任务开工前**插入一道需求澄清环节：
@@ -18,6 +58,8 @@ AI 助手最常见的失败不是"不会做"，而是"做完了才发现做的
 
 严格档位（`gateMode: "strict"`）下，需求没确认前模型无法执行写文件、跑命令这类**变更类**操作——从机制上保证"没问清就不许动手"。
 
+`ask` 档位（0.1.4 新增）介于两者之间：需求没确认时，变更类操作会弹一张 DSH 原生确认卡片问你"照现在的理解直接动手吗"，你点「是」就放行，点「否」由你补一句需求——不写配置文件也能当场决定。
+
 ## 安装
 
 装插件前请先退出正在跑的任务。三种方式任选：
@@ -35,7 +77,7 @@ dsh plugin --profile <你的 profile 名> add dsh-grill-first
 **方式三：本地包 / GitHub 仓库**
 
 ```bash
-dsh plugin --profile <你的 profile 名> add /绝对路径/dsh-grill-first-0.1.3.tgz
+dsh plugin --profile <你的 profile 名> add /绝对路径/dsh-grill-first-0.1.5.tgz
 dsh plugin --profile <你的 profile 名> add github:InoriMutsumi/dsh-grill-first
 ```
 
@@ -56,38 +98,41 @@ dsh plugin --profile <你的 profile 名> add github:InoriMutsumi/dsh-grill-firs
 ### 想跳过提问时
 
 - 消息里带上「**直接做**」「**跳过提问**」「**不用问**」「**skip**」中任意一个 → 立刻开工；
+- 只是**问情况**（"这个插件怎么样""查一下我的仓库里有什么""解释一下这段代码"）→ 不算新任务，不追问；
 - 或者在设置里把 `enabled` 关掉 / 把 `gateMode` 调成 `off` → 整个插件静默。
 
 ## 配置项
 
-DSH 设置界面 → 插件 → dsh-grill-first，共 7 项：
+DSH 设置界面 → 插件 → dsh-grill-first，共 8 项：
 
 | 配置 | 类型 | 默认 | 说明 |
 |---|---|---|---|
 | `enabled` | 布尔 | `true` | 总开关。关闭后既不追问也不拦截工具 |
-| `gateMode` | `off` / `warn` / `strict` | `warn` | `off` 不拦截；`warn` 只在日志里记录"需求未确认"；`strict` 未确认前拦截变更类工具 |
+| `gateMode` | `off` / `warn` / `ask` / `strict` | `warn` | `off` 不拦截；`warn` 只在日志里记录"需求未确认"；`ask` 未确认时弹卡片问你；`strict` 未确认前直接拦截变更类工具 |
 | `maxQuestionsPerRound` | 1-6 | `3` | 每轮最多问几个问题 |
-| `maxRounds` | 1-10 | `3` | 最多追问几轮 |
-| `intensity` | `gentle` / `standard` / `grill` | `standard` | 追问力度：只问关键点 / 标准 / 刨根问底 |
+| `maxRounds` | 1-10 | `3` | 最多追问几轮；问满后 `grill_start` 直接放行 |
+| `intensity` | `gentle` / `standard` / `grill` | `standard` | 追问力度：`gentle` 两维齐就放行 / `standard` 三维 / `grill` 还要求写清验收标准 |
 | `bypassKeywords` | 字符串数组 | `["直接做","跳过提问","不用问","skip"]` | 命中即跳过提问 |
 | `memoryEnabled` | 布尔 | `true` | 是否启用两层记忆 |
+| `memoryScope` | `project` / `global` | `project` | 项目背景/历史存哪：`project` 存工作区，`global` 存全局目录（跨项目共享） |
 
 > 默认是 `warn` 而不是 `strict`：默认档位保证升级插件不会突然卡住你已有的工作流。想要硬门禁请手动切 `strict`。
 
 ### 怎么改配置
 
-DSH 设置页里的插件条目目前只提供「启用 / 停用」开关（客户端的插件页不渲染第三方插件的配置表单），要改上面 7 项请编辑 **profile 的补丁层** `cordis.patch.yml`：
+DSH 设置页里的插件条目目前只提供「启用 / 停用」开关（客户端的插件页不渲染第三方插件的配置表单），要改上面 8 项请编辑 **profile 的补丁层** `cordis.patch.yml`：
 
 ```yaml
 # <profile 目录>/cordis.patch.yml
 - id: dsh-grill-first
   name: dsh-grill-first
   config:
-    gateMode: strict          # off | warn | strict
+    gateMode: strict          # off | warn | ask | strict
     maxQuestionsPerRound: 3   # 1-6
     maxRounds: 3              # 1-10
     intensity: standard       # gentle | standard | grill
     memoryEnabled: true
+    memoryScope: project      # project | global
     bypassKeywords: [直接做, 跳过提问, 不用问, skip]
 ```
 
@@ -102,6 +147,8 @@ DSH 设置页里的插件条目目前只提供「启用 / 停用」开关（客
 | 全局层 | `<profile 目录>/grill-first/global.json`（不可写时退化到 `~/.dsh-grill-first/global.json`） | 通用偏好，如"注释一律用中文" |
 | 项目层 | `<工作区>/.dsh/grill-first/memory.json` | 项目背景、任务历史（任务名 + 需求摘要 + 日期） |
 
+项目层默认落在工作区；把 `memoryScope` 改成 `global` 后改落 `<全局目录>/project.json`——适合"所有项目通用"的背景约定，不想往每个仓库里塞 `.dsh/` 时也用这个。历史记录只保留最近 100 条（0.1.4 起）。
+
 下次追问时，`grill_start` 会把这些内容作为 `memory_context` 返回，模型就不会重复问你同一件事。
 
 ## 开发
@@ -109,7 +156,7 @@ DSH 设置页里的插件条目目前只提供「启用 / 停用」开关（客
 ```bash
 npm install          # 装开发依赖（schemastery 由宿主提供，仅开发期需要）
 npm run build        # TypeScript → dist/
-npm test             # 构建 + 94 个单测 / 端到端场景测试
+npm test             # 构建 + 127 个单测 / 端到端场景测试
 node scripts/selfcheck.mjs   # 静态自检：中文 JSDoc、any 用量、需求→代码位置
 ```
 
@@ -121,7 +168,7 @@ node scripts/selfcheck.mjs   # 静态自检：中文 JSDoc、any 用量、需求
 src/
   index.ts              插件装配入口（工具 + 协议段 + 门禁）
   config.ts             配置归一化（不依赖宿主，可单测）
-  configSchema.ts       Schemastery 配置声明（设置界面用，7 项）
+  configSchema.ts       Schemastery 配置声明（设置界面用，8 项）
   version.ts            从 package.json 读版本号，供 grill_start 返回
   judge/heuristics.ts   任务模糊度判定（纯函数）
   prompt/protocol.ts    注入系统提示的中文协议文本
@@ -145,7 +192,7 @@ src/
 > DSH 自己的门禁用 `semver.satisfies(..., { includePrerelease: true })` 判定，`^0.2.0-rc.2` 能匹配 `0.2.1-rc.2`；
 > 但标准 npm 语义（不加 `includePrerelease`）会判不匹配，所以上游要求写显式分支。
 
-已验证的宿主版本：`0.2.0-rc.2`（0.1.x 开发期）、`0.2.1-rc.2`（2026-10-17 复核：`typecheck`、94 个单测、`verify-installed` 一致，且实测加载运行）。升完 DSH 后的自检三步：
+已验证的宿主版本：`0.2.0-rc.2`（0.1.x 开发期）、`0.2.1-rc.2`（2026-10-17 复核：`typecheck`、94 个单测、`verify-installed` 一致，且实测加载运行；0.1.4 起 127 个单测）。升完 DSH 后的自检三步：
 
 ```bash
 node scripts/verify-installed.mjs   # 安装拷贝与仓库产物是否一致
@@ -156,6 +203,8 @@ node scripts/verify-installed.mjs   # 安装拷贝与仓库产物是否一致
 ## 已知限制
 
 - 门禁只对**根 agent**（能直接跟你对话的那个）生效；子 agent / 团队成员不拦，否则正常的委派工作流会被卡死。因此**委派类工具本身**（`subagent`、`workflow`、`spawn_teammate`…）在未确认前也会被拦，防止"派个子代理替自己动手"绕过门禁。
+- `ask` 档位**只问一次**：卡片弹过之后门禁即视为放行（否则一次任务里写 20 个文件就要弹 20 张卡片）。想每次都拦，请用 `strict`。
+- 判定器是**关键词启发式**，不是语义模型：问句靠问号/疑问词/"查一下/看看"这类信号识别，极少数既像提问又像下令的句子可能判错；判错时直接说「直接做」或让模型调一次 `grill_start` 即可纠正。英文关键词按词边界匹配，中文按子串匹配。
 - `warn` 档位只能"记录不拦截"：DSH 的 `tools/pre-execute` 事件没有"放行但同时给模型提示"这种返回形态。
 - 记忆是本地文件，不含加密；删掉对应 json 即清空。
 - `grill_confirm` 优先走 DSH 的结构化提问通道；该通道不可用时（例如子 agent 内部）会返回 `needs_user_confirmation: true`，由模型改用内置提问工具兜底。

package/cordis.patch.yml

@@ -1,4 +1,4 @@
-# DSH bundle 补丁：把本插件插入当前 profile 的插件表（格式参照已装插件 dsh-find-plugins）。
-- insert:
-    - id: dsh-grill-first
-      name: 'dsh-grill-first'
+# DSH bundle 补丁：把本插件插入当前 profile 的插件表（格式参照已装插件 dsh-find-plugins）。
+- insert:
+    - id: dsh-grill-first
+      name: 'dsh-grill-first'

package/dist/config.d.ts

@@ -1,5 +1,7 @@
-/** 门禁档位：off=不拦截、warn=只记录、strict=拦截变更类工具。 */
-export type GateMode = 'off' | 'warn' | 'strict';
+/** 门禁档位：off=不拦截、warn=只记录、strict=拦截变更类工具、ask=先弹卡片问用户。 */
+export type GateMode = 'off' | 'warn' | 'strict' | 'ask';
+/** 项目层记忆的落点：project=跟着当前工作区走、global=全局共享一份。 */
+export type MemoryScope = 'project' | 'global';
 /** 追问力度。 */
 export type Intensity = 'gentle' | 'standard' | 'grill';
 /** 归一化后的插件配置。 */
@@ -18,6 +20,8 @@ export interface GrillConfig {
     bypassKeywords: string[];
     /** 是否启用两层记忆（R5）。 */
     memoryEnabled: boolean;
+    /** 项目层记忆落在哪里（0.1.4 新增：有些项目不想被写进 .dsh/）。 */
+    memoryScope: MemoryScope;
 }
 /** 默认配置（与工作表 3.7 的默认值一致）。 */
 export declare const DEFAULT_CONFIG: GrillConfig;

package/dist/config.js

@@ -14,6 +14,7 @@ export const DEFAULT_CONFIG = {
     intensity: 'standard',
     bypassKeywords: [...DEFAULT_BYPASS_KEYWORDS],
     memoryEnabled: true,
+    memoryScope: 'project',
 };
 /** 取整并夹到 [min, max]；非法值返回 fallback。 */
 function clampInt(value, min, max, fallback) {
@@ -53,11 +54,12 @@ export function normalizeConfig(raw) {
     const input = (raw ?? {});
     return {
         enabled: pickBoolean(input.enabled, DEFAULT_CONFIG.enabled),
-        gateMode: pickEnum(input.gateMode, ['off', 'warn', 'strict'], DEFAULT_CONFIG.gateMode),
+        gateMode: pickEnum(input.gateMode, ['off', 'warn', 'strict', 'ask'], DEFAULT_CONFIG.gateMode),
         maxQuestionsPerRound: clampInt(input.maxQuestionsPerRound, 1, 6, DEFAULT_CONFIG.maxQuestionsPerRound),
         maxRounds: clampInt(input.maxRounds, 1, 10, DEFAULT_CONFIG.maxRounds),
         intensity: pickEnum(input.intensity, ['gentle', 'standard', 'grill'], DEFAULT_CONFIG.intensity),
         bypassKeywords: pickStringList(input.bypassKeywords, DEFAULT_CONFIG.bypassKeywords),
         memoryEnabled: pickBoolean(input.memoryEnabled, DEFAULT_CONFIG.memoryEnabled),
+        memoryScope: pickEnum(input.memoryScope, ['project', 'global'], DEFAULT_CONFIG.memoryScope),
     };
 }

package/dist/configSchema.d.ts

@@ -10,18 +10,20 @@ import z from '@deepseek-ai/schemastery';
 /** 插件配置 schema：改了这里，设置界面与 cordis.yml 校验同时生效。 */
 export declare const Config: z<Schemastery.ObjectS<NoInfer<{
     enabled: z<boolean, boolean, "defined">;
-    gateMode: z<"off" | "warn" | "strict", "off" | "warn" | "strict", "defined">;
+    gateMode: z<"off" | "warn" | "strict" | "ask", "off" | "warn" | "strict" | "ask", "defined">;
     maxQuestionsPerRound: z<number, number, "defined">;
     maxRounds: z<number, number, "defined">;
     intensity: z<"gentle" | "standard" | "grill", "gentle" | "standard" | "grill", "defined">;
     bypassKeywords: z<string[], string[], "defined">;
     memoryEnabled: z<boolean, boolean, "defined">;
+    memoryScope: z<"project" | "global", "project" | "global", "defined">;
 }>>, Schemastery.ObjectT<NoInfer<{
     enabled: z<boolean, boolean, "defined">;
-    gateMode: z<"off" | "warn" | "strict", "off" | "warn" | "strict", "defined">;
+    gateMode: z<"off" | "warn" | "strict" | "ask", "off" | "warn" | "strict" | "ask", "defined">;
     maxQuestionsPerRound: z<number, number, "defined">;
     maxRounds: z<number, number, "defined">;
     intensity: z<"gentle" | "standard" | "grill", "gentle" | "standard" | "grill", "defined">;
     bypassKeywords: z<string[], string[], "defined">;
     memoryEnabled: z<boolean, boolean, "defined">;
+    memoryScope: z<"project" | "global", "project" | "global", "defined">;
 }>>, "plain">;

package/dist/configSchema.js

@@ -9,13 +9,15 @@
 import z from '@deepseek-ai/schemastery';
 import { DEFAULT_BYPASS_KEYWORDS } from './judge/heuristics.js';
 /** 门禁档位 schema。 */
-const gateModeSchema = z.union([z.const('off'), z.const('warn'), z.const('strict')]).default('warn');
+const gateModeSchema = z
+    .union([z.const('off'), z.const('warn'), z.const('strict'), z.const('ask')])
+    .default('warn');
 /** 追问力度 schema。 */
 const intensitySchema = z.union([z.const('gentle'), z.const('standard'), z.const('grill')]).default('standard');
 /** 插件配置 schema：改了这里，设置界面与 cordis.yml 校验同时生效。 */
 export const Config = z.object({
     enabled: z.boolean().default(true).description('总开关：关闭后任何任务都不再追问，也不拦截工具。'),
-    gateMode: gateModeSchema.description('门禁档位：off 不拦截 / warn 只记录 / strict 未确认需求前拦截写类工具。'),
+    gateMode: gateModeSchema.description('门禁档位：off 不拦截 / warn 只记录 / strict 未确认需求前拦截写类工具 / ask 拦截前先弹卡片问一句。'),
     maxQuestionsPerRound: z
         .natural()
         .max(6)
@@ -28,4 +30,8 @@ export const Config = z.object({
         .default([...DEFAULT_BYPASS_KEYWORDS])
         .description('跳过关键词：用户消息里出现这些词就不再追问，直接开工。'),
     memoryEnabled: z.boolean().default(true).description('是否启用两层记忆（全局偏好 + 项目背景/历史）。'),
+    memoryScope: z
+        .union([z.const('project'), z.const('global')])
+        .default('project')
+        .description('项目层记忆落点：project 写进当前工作区的 .dsh/grill-first/，global 写进插件家目录。'),
 });

package/dist/gate/preExecute.d.ts

@@ -9,8 +9,8 @@
  * 3. warn 模式只记录不拦截（pre-execute 没有"放行 + 附带提示"这种返回形态）。
  */
 import type { GrillContext } from '../types.js';
-/** 门禁档位。 */
-export type GateMode = 'off' | 'warn' | 'strict';
+/** 门禁档位：ask 档在拦截前先弹一张"要不要继续"的卡片，交给人决定（0.1.4 新增）。 */
+export type GateMode = 'off' | 'warn' | 'strict' | 'ask';
 /** 门禁状态：open = 已确认或无需确认；closed = 尚未确认。 */
 export type GateState = 'open' | 'closed';
 /** 变更类工具名单（宁松勿紧）。 */
@@ -19,14 +19,16 @@ export declare const MUTATING_TOOLS: readonly string[];
 export declare const READ_ONLY_TOOLS: readonly string[];
 /** 拦截时给模型看的拒绝文案（工作表 3.6 原文）。 */
 export declare const GATE_DENY_REASON = "\u9700\u6C42\u5C1A\u672A\u786E\u8BA4\uFF1A\u8BF7\u5148\u8C03\u7528 grill_confirm \u5B8C\u6210\u9700\u6C42\u786E\u8BA4\uFF0C\u6216\u8BA9\u7528\u6237\u8BF4'\u76F4\u63A5\u505A'\u8DF3\u8FC7\u3002";
+/** ask 档弹给用户看的说明（宿主用 displayReason 渲染卡片正文）。 */
+export declare const GATE_ASK_REASON = "\u8FD9\u6B21\u4EFB\u52A1\u7684\u9700\u6C42\u8FD8\u6CA1\u786E\u8BA4\uFF0C\u8981\u7167\u73B0\u5728\u7684\u7406\u89E3\u76F4\u63A5\u52A8\u624B\u5417\uFF1F";
 /**
  * 判断某个工具名是否属于"会真正改动工作区/交付物"的变更类工具。
  * @param name 工具名。
  */
 export declare function isMutatingTool(name: string): boolean;
-/** 门禁判定结果。 */
+/** 门禁判定结果：ask = 弹卡片问用户后由其决定（宿主 PreToolDecision 支持）。 */
 export interface GateDecision {
-    action: 'allow' | 'deny';
+    action: 'allow' | 'deny' | 'ask';
     reason?: string;
 }
 /**
@@ -60,6 +62,14 @@ export interface GateOptions {
     mode: GateMode;
     /** warn 模式下的提示回调（默认写日志）。 */
     onWarning?: (toolName: string, agentId: string) => void;
+    /**
+     * 新用户消息到达时，是否把门禁重置为 closed。
+     *
+     * 0.1.4 修：以前每条用户消息都关门禁，导致"刚确认完、用户回一句继续"又被拦。
+     * 现在由判定器看一眼消息内容——追问/询问/说了跳过词之类不关门禁。
+     * 返回 true 才关门；未提供该回调时保持旧行为（一律关）。
+     */
+    shouldResetOnMessage?: (text: string) => boolean;
 }
 /**
  * 把门禁挂到宿主事件上：`tools/pre-execute` 拦截 + `agent/inbox/inserted` 重置。