Pi skips sub-agents and plan mode.
This harness builds them.Pi 刻意不做子代理和计划模式。
这个 harness 把它们全部补上。

Install安装

One command, zero companion extensions. Restart pi and you’re done.一条命令，无需任何伴随扩展。重启 pi 即可。

pi install npm:pi-muselinn-harness

Or from source: pi install git:github.com/MuseLinn/pi-muselinn-harness或源码安装：pi install git:github.com/MuseLinn/pi-muselinn-harness

Try it体验

Everything is a slash command or a model-callable tool — no config files to learn first.一切都是 slash 命令或模型可调用工具——无需先学习任何配置文件。

/swarm on                                # parallel sub-agents
/goal Refactor the auth module           # budget-tracked goal
/todo init "Phase 1: scanner"            # phased task plan
/plan                                    # read-only plan mode
/pause                                   # freeze; esc/enter/space/ctrl+c resumes
/tui style plain|boxed|compact           # switch editor chrome

Docs文档

Full feature walkthroughs, command/tool references, architecture, and tests live in the README (EN / 中).完整功能说明、命令/工具参考、架构与测试详见 README（EN / 中）。

• README (EN)
• README (中文)
• Changelog · Self-hosting

Type-layer hardening — full-package typecheck类型层加固 — 全包类型检查

The package now ships tsconfig.json + npm run typecheck (strict, es2024) wired into CI, clearing 158 type errors against pi 0.83 — and the new checker surfaced real runtime bugs: resources_discover injected subagent skills through an unimported function (the ReferenceError was swallowed by catch, so skills were silently missing), and todo cleanup was bound to a session_end event that pi 0.83 never emits (now session_shutdown). Dead usage fields are gone from every command definition, vision-model detection uses pi’s real Model type, and notify(..., "success") is typed via module augmentation.本包现在自带 tsconfig.json + npm run typecheck（strict, es2024）并接入 CI，清除了针对 pi 0.83 的 158 条类型错误——新检查器还暴露了真实运行时缺陷：resources_discover 通过未导入的函数注入子代理技能（ReferenceError 被 catch 静默吞掉，技能一直缺失）、todo 清理绑定到 pi 0.83 从不触发的 session_end 事件（现已改为 session_shutdown）。所有命令定义中的死字段 usage 已删除，视觉模型检测改用 pi 的真实 Model 类型，notify(..., "success") 通过模块增强补齐类型。

RPC approvals & window-aware truncationRPC 审批与窗口感知截断

Approval dialogs now work in RPC hosts (obsidian-pi & other embedding clients): the same Allow once / Always allow / Deny / Deny with reason choices render over the extension UI protocol (select / input / confirm) instead of silently denying every edit (issue #4). The tool-result spill threshold now scales with the model’s context window (max(40k, window × 4 chars/token), capped at 800k chars ≈ 200k tokens), with PI_TRUNCATION_THRESHOLD as an explicit override — 1M-context models keep far more output in-context (issue #2).审批弹窗在 RPC 宿主（obsidian-pi 等嵌入客户端）中可用：允许一次 / 本会话始终允许 / 拒绝 / 拒绝并附原因，通过扩展 UI 协议（select / input / confirm）呈现，不再静默拒绝每次编辑（issue #4）。工具结果落盘阈值随模型上下文窗口缩放（max(40k, 窗口 × 4 字符/token)，上限 800k 字符 ≈ 200k token），可用 PI_TRUNCATION_THRESHOLD 显式覆盖——1M 上下文模型可保留远多输出（issue #2）。

Package hotfix — pause/ shipped打包修复 — pause/ 随包发布

The published tarball was missing the pause/ adapter directory and the extension failed to load (Cannot find module './pause/commands'); it is now part of the package (issue #3).发布的 tarball 缺少 pause/ 适配目录导致扩展加载失败（Cannot find module './pause/commands'）；现已纳入发布包（issue #3）。

Freeze & steer — /pause · transcript · /steer冻结与转向 — /pause · transcript · /steer

/pause parks the main agent and every swarm subagent at their next safe boundary (tool-call gate): in-flight calls finish, nothing is aborted, release continues exactly where each loop parked — behind a full-screen overlay with a theme-colored pause glyph and a live hold timer (esc/enter/space/ctrl+c). Every subagent’s conversation is recorded to <sessionDir>/agents/<id>/wire.jsonl (no tool arguments), viewable with c in the task browser. /steer <taskId> <message> injects a message into a running subagent (swarm session or background task); the loop delivers it after the current tool call. Releasing leaves a status line in the session transcript./pause 把主代理与所有 swarm 子代理停在下一次安全边界（工具门卫）：进行中的调用跑完、什么都不中止、释放后从挂起点继续——全屏遮罩,主题色暂停符号 + 实时计时（esc/enter/space/ctrl+c）。每个子代理的对话落盘 <sessionDir>/agents/<id>/wire.jsonl（不落工具参数）,任务浏览器按 c 查看。/steer <taskId> <message> 向运行中的子代理注入消息（swarm 会话或后台任务）,循环在当前工具调用后投递。释放时在会话流中留下状态行。

0.9.18 — Persistent startup permission mode0.9.18 — 可持久化的启动权限模式

Set "defaultMode": "auto" | "yolo" | "manual" in permissions.json (global ~/.pi/agent/ or project .pi/, global wins on conflict) and every fresh session starts there — no more interactive /mode on each new session. Sessions with a recorded /mode history still restore the last used mode; defaultMode is the starting point for fresh sessions. Thanks @jason-in-hub for the contribution!在 permissions.json（全局 ~/.pi/agent/ 或项目 .pi/，冲突时全局优先）写入 "defaultMode": "auto" | "yolo" | "manual"，每个新会话都从该模式开始——不再需要每次手动 /mode。记录过 /mode 历史的会话仍恢复上次使用的模式；defaultMode 是新会话的起始点。感谢 @jason-in-hub 的贡献！