emrg: the READMEs read their facts from the code, and a guard keeps them true

Agent.md

@@ -34,7 +34,7 @@ The execution model, in this repo's own terms (`emrg/server/daemon.py`):
 
 - **Tool loop** — the whole process triggered by one user message: LLM requests and tool calls repeat until a round produces no new tool calls (`_run_tool_loop`).
 - **Round** — one iteration inside a tool loop: one LLM request + zero or more tool calls + their execution (`round_num`, bounded by `max_tool_rounds`).
-- **Evolution cycle** — one full self-evolution run (Prepare → Review → Discover → Improve → Submit → Record). Never call it a "round".
+- **Evolution cycle** — one full self-evolution run (Preparation → Look → Think → Act → Record). Never call it a "round".
 
 ## Capabilities
 

README.cn.md

@@ -47,7 +47,7 @@ EMRG 是一个关于*自主进化*的实验：一个能帮你写代码的 AI 智
 |---|---|
 | **Electron GUI（主入口）** | 安装即用：首次启动引导配置 API Key；全部 `/` 指令可用；结果面板、Ask/Auto 模式、自进化可见化。GUI 配好即 TUI 可用 |
 | **全功能 TUI + 守护进程** | 流式 Markdown、`/` 自动补全、会话管理、快捷键——跑在常驻的 `emrgd` 守护进程上，随时重连 |
-| **定时任务** | 内置 + 自定义任务类型（提示词模板存 `~/.emrg/task-templates/`）、GUI 设置页增删改 + 热加载、`/trigger` 随时手动触发 |
+| **定时任务** | 内置任务类型（`evolution`、`paper`、`open-source`、`promote`、`journal`、`competition`、`ios-app`、`feishu`）加自定义类型（提示词模板存 `~/.emrg/task-templates/`）、GUI 设置页增删改 + 热加载、`/trigger` 随时手动触发 |
 | **100% 开源** | MIT 协议——没有围墙，没有厂商锁定。默认英文，提供中文版 |
 
 ---
@@ -64,8 +64,8 @@ EMRG 不只是一个工具——它是一个**会听吐槽、会自我改进**
   - 跨项目学习
           |
           v
-演化循环（每 30 分钟）
-（准备 -> 回顾 -> 发现 -> 改进 -> 提交 -> 记录）
+演化循环（默认每 10 分钟）
+（Preparation -> Look -> Think -> Act -> Record）
           |
           v
   1. pytest + import 检查
@@ -148,12 +148,14 @@ TUI 内：`Esc` 中断流式输出，`Ctrl+C` / `exit` 退出（配置与 GUI 
 | `/memory` | 浏览项目和会话记忆 |
 | `/rename [标题]` | 给当前会话起个好记的名字 |
 | `/model [name]` | 切换 LLM 模型——不带参数进入交互式选择器 |
+| `/sandbox [mode]` | 设置本会话的沙箱档位——不带参数弹出选择器；`add`/`remove`/`list` 管理额外可写根 |
 | `/rant <反馈> [@<project>]` | 吐槽、建议、夸奖——演化系统会听 |
 | `/help` | 查看所有键盘快捷键和命令帮助 |
 | `/image` | 从剪贴板插入图片到输入框 |
 | `/delete [id]` | 删除会话——不带参数进入交互式选择器 |
 | `/rewind` | 回退对话到历史某个节点 |
 | `/trigger` | 触发演化任务——交互式选择器（任务管理在 GUI 设置 → 定时任务） |
+| `/task-session` | 打开某个定时任务自己的会话——交互式选择器 |
 | `/skills` | 列出已加载技能；`/skills available\|install\|update` |
 | `/version` | 显示 EMRG 版本和实例信息 |
 | `Esc` | 中断正在运行的响应 |
@@ -166,13 +168,13 @@ TUI 内：`Esc` 中断流式输出，`Ctrl+C` / `exit` 退出（配置与 GUI 
 |  | Claude Code | Codex | DeepSeek Harness | **EMRG** | **Grok Bot** |
 |---|---|---|---|---|---|
 | AI 驱动编程 | ✅ | ✅ | ✅ | ✅ | ✅ |
-| 工具调用 (bash, read, write, edit, glob, grep) | ✅ | ✅ | ✅ | ✅ | ✅ |
+| 工具调用 (`bash`/`pwsh`、`read`、`write`、`edit`、`glob`、`grep`、`submit_rant`、`subagent`) | ✅ | ✅ | ✅ | ✅ | ✅ |
 | 会话记忆与上下文 | ✅ | ✅ | ✅ | ✅ | ✅ |
 | `/` 命令自动补全 | ✅ | ✅ | ✅ | ✅ | ~ *@ mention 建议 + 命令面板* |
 | ESC 中断 | ✅ | ✅ | — | ✅ | ✅ |
 | 插件化架构 | ❌ | ❌ | ✅ *一切皆插件* | ❌ | ✅ |
 | 沙箱执行 | ❌ | ❌ | ✅ | ✅ *read-only / workspace-write / danger-full-access* | ~ *protectedBoxPaths + /workspace* |
-| 子代理 | ❌ | ❌ | ✅ | ❌ | ✅ *assistants/automations 生态* |
+| 子代理 | ❌ | ❌ | ✅ | ✅ *子会话委派* | ✅ *assistants/automations 生态* |
 | Web UI | ❌ | ❌ | ✅ *浏览器* | ✅ *Electron* | ✅ *Electron* |
 | **递归自我改进（RSI）** | ❌ | ❌ | ❌ | ✅ *全自动——代码、提示词、工具链* | ❌ |
 | **个人智能体** *你的机器、你的 key、你的数据* | ~ *本地 CLI，模型为托管服务，走你的账号* | ~ *本地 CLI，模型为托管服务* | ~ *本地 harness，模型为托管服务* | ✅ *本地智能体，用你的 key，状态在你磁盘上* | ~ *本地应用，托管服务* |

README.md

@@ -47,7 +47,7 @@ EMRG is an experiment in *autonomous self-improvement*: an AI agent that helps y
 |---|---|
 | **Electron GUI (main entry)** | Install & go: first-run wizard, all slash commands, results panel, Ask/Auto modes, visible self-evolution. GUI configured = TUI ready |
 | **Full-featured TUI + daemon** | Streaming markdown, `/` autocomplete, sessions, shortcuts — on a persistent `emrgd` daemon you can reconnect to anytime |
-| **Scheduled tasks** | Built-in + custom task types with prompt templates (`~/.emrg/task-templates/`), CRUD + hot reload from the GUI settings, `/trigger` any task on demand |
+| **Scheduled tasks** | Built-in task types (`evolution`, `paper`, `open-source`, `promote`, `journal`, `competition`, `ios-app`, `feishu`) plus custom ones with prompt templates in `~/.emrg/task-templates/`, CRUD + hot reload from the GUI settings, `/trigger` any task on demand |
 | **100% open source** | MIT — no walled garden, no vendor lock-in. English default, Chinese version available |
 
 ---
@@ -64,8 +64,8 @@ Inputs:
   - Cross-project learning
           |
           v
-Evolution Cycle (every 30 min)
-(Prepare -> Review -> Discover -> Improve -> Commit -> Record)
+Evolution Cycle (every 10 min by default)
+(Preparation -> Look -> Think -> Act -> Record)
           |
           v
   1. pytest + import check
@@ -150,12 +150,14 @@ In the TUI: `Esc` interrupts streaming, `Ctrl+C` / `exit` quits (config is share
 | `/memory` | Browse project & session memories |
 | `/rename [title]` | Give your session a memorable name |
 | `/model [name]` | Switch LLM model — no args for interactive picker |
+| `/sandbox [mode]` | Set this session's sandbox tier — no args for a picker; `add`/`remove`/`list` manage extra writable roots |
 | `/rant <feedback> [@<project>]` | Complain, suggest, praise — evolution listens |
 | `/help` | Show keyboard shortcuts and command help |
 | `/image` | Insert clipboard image into the input field |
 | `/delete [id]` | Delete a session — no args for interactive picker |
 | `/rewind` | Rewind conversation to a history point |
 | `/trigger` | Trigger an evolution task — interactive picker (manage tasks in GUI Settings → Scheduled tasks) |
+| `/task-session` | Open a scheduled task's own session — interactive picker |
 | `/skills` | List loaded skills; `/skills available\|install\|update` |
 | `/version` | Show EMRG version and instance info |
 | `Esc` | Interrupt a running response mid-stream |
@@ -168,13 +170,13 @@ In the TUI: `Esc` interrupts streaming, `Ctrl+C` / `exit` quits (config is share
 |  | Claude Code | Codex | DeepSeek Harness | **EMRG** | **Grok Bot** |
 |---|---|---|---|---|---|
 | AI-powered coding | ✅ | ✅ | ✅ | ✅ | ✅ |
-| Tool-calling (bash, read, write, edit, glob, grep) | ✅ | ✅ | ✅ | ✅ | ✅ |
+| Tool-calling (`bash`/`pwsh`, `read`, `write`, `edit`, `glob`, `grep`, `submit_rant`, `subagent`) | ✅ | ✅ | ✅ | ✅ | ✅ |
 | Session memory & context | ✅ | ✅ | ✅ | ✅ | ✅ |
 | `/` command autocomplete | ✅ | ✅ | ✅ | ✅ | ~ *@ mention + command palette* |
 | ESC interrupt | ✅ | ✅ | — | ✅ | ✅ |
 | Plugin architecture | ❌ | ❌ | ✅ *everything is a plugin* | ❌ | ✅ |
 | Sandboxed execution | ❌ | ❌ | ✅ | ✅ *read-only / workspace-write / danger-full-access* | ~ *protectedBoxPaths + /workspace* |
-| Subagents | ❌ | ❌ | ✅ | ❌ | ✅ *assistants/automations* |
+| Subagents | ❌ | ❌ | ✅ | ✅ *delegated child sessions* | ✅ *assistants/automations* |
 | Web UI | ❌ | ❌ | ✅ *browser* | ✅ *Electron* | ✅ *Electron* |
 | **Recursive self-improvement (RSI)** | ❌ | ❌ | ❌ | ✅ *autonomous — code, prompts, toolchain* | ❌ |
 | **Personal agent** *your machine, your key, your data* | ~ *local CLI, hosted model, your account* | ~ *local CLI, hosted model* | ~ *local harness, hosted model* | ✅ *local agent, your key, its state on your disk* | ~ *local app, hosted service* |

emrg/server/daemon.py

@@ -639,6 +639,50 @@ def build_shell_tool(
     return BashToolV2()
 
 
+def build_tool_registry(
+    sandbox_config: Optional[SandboxConfig] = None,
+    run_subagent_cb=None,
+    platform_name: str | None = None,
+) -> ToolRegistry:
+    """The daemon's tool set, built in one place so that its size is a reading.
+
+    The registry is the one fact about this agent that a document cannot restate
+    by hand and stay true: `README.md` names the tools, and a name added here is a
+    name that document stops being right about the moment it is added. Something
+    therefore has to compare the two, and a comparison that lives in a test needs a
+    producer it can call **without a daemon** — a copy of this block in the test
+    would be a second list to keep in step, which is the defect
+    `emrg/tools/shell_dialects.py` exists to name. So the block is here, the daemon
+    calls it, and `tests/test_readme_matches_code.py` reads the same function.
+
+    :param sandbox_config: the ``[sandbox]`` section, passed through to
+        :func:`build_shell_tool`; a missing one means the defaults.
+    :param run_subagent_cb: the caller's own runner. The tool has no path to the
+        session machinery of its own, so everything that owns state stays with the
+        daemon that supplies it.
+    :param platform_name: the platform to decide the shell dialect for; defaults to
+        the host's, which is what the daemon wants and what a test overrides.
+    :returns: the populated registry.
+    """
+    tools = ToolRegistry()
+    # Which shell this platform mounts — and, on POSIX, which of the two
+    # executors answers to the name `bash`.  The gate is composition data
+    # rather than a runtime branch (design §14.2 layer 6): Windows has no
+    # `bash` tool row at all, so "bash is not found" is not representable
+    # there.  See `build_shell_tool` for the three cases.
+    tools.register(build_shell_tool(sandbox_config, platform_name))
+    tools.register(ReadTool())
+    tools.register(WriteTool())
+    tools.register(EditTool())
+    tools.register(GlobTool())
+    tools.register(GrepTool())
+    tools.register(SubmitRantTool())
+    # The runner is the daemon's own method, so the tool has no path to the
+    # session machinery of its own: everything that owns state stays there.
+    tools.register(SubagentTool(run_subagent_cb))
+    return tools
+
+
 class EmrgServer:
     """EMRG daemon — listens on Unix socket, processes tasks with tool calling."""
 
@@ -798,22 +842,7 @@ def __init__(self, llm_config: LlmConfig) -> None:
         self._pending_web_auth_proc: Optional[asyncio.subprocess.Process] = None
 
         # Build tool registry
-        self.tools = ToolRegistry()
-        # Which shell this platform mounts — and, on POSIX, which of the two
-        # executors answers to the name `bash`.  The gate is composition data
-        # rather than a runtime branch (design §14.2 layer 6): Windows has no
-        # `bash` tool row at all, so "bash is not found" is not representable
-        # there.  See `build_shell_tool` for the three cases.
-        self.tools.register(build_shell_tool(self._sandbox_config))
-        self.tools.register(ReadTool())
-        self.tools.register(WriteTool())
-        self.tools.register(EditTool())
-        self.tools.register(GlobTool())
-        self.tools.register(GrepTool())
-        self.tools.register(SubmitRantTool())
-        # The runner is the daemon's own method, so the tool has no path to the
-        # session machinery of its own: everything that owns state stays here.
-        self.tools.register(SubagentTool(self._run_subagent_tool))
+        self.tools = build_tool_registry(self._sandbox_config, self._run_subagent_tool)
         logger.info("tools registered: %s", self.tools.names)
 
         # Skills are **not** held here. A snapshot taken at startup cannot see a skill