@gavin-hu/my-pi-agent 0.16.0 → 0.18.0

package/README.md

@@ -36,7 +36,8 @@ want the agent to *think first and work safely*:
   plan file that the agent then follows step by step.
 - **Delegate and parallelize.** Run specialist subagents in their own Pi
   processes, push a long build into a background job, and let a watcher tail a
-  log until it matters — all while you keep the thread.
+  log until it matters — then run several sessions on one repository through a
+  shared bus, each owning its slice, all while you keep the thread.
 - **Never wreck the checkout.** Isolation is one command: the task moves onto its
   own branch in a `git worktree`, and every prompt is snapshotted so a bad turn
   is one `/rewind` away.
@@ -47,7 +48,7 @@ want the agent to *think first and work safely*:
   through a bridge to the live session.
 
 Everything is optional and independently toggleable. Nothing is forked, and no
-installed Pi file is modified: Pi discovers the resources through the `pi`
+installed Pi file is written to: Pi discovers the resources through the `pi`
 manifest in [`package.json`](./package.json), so the package upgrades with Pi
 like any other.
 
@@ -74,34 +75,40 @@ Then, from inside Pi:
   every loaded extension with its runtime health, enables or disables one, and
   clears a project override with `x`.
 - `/bytefly settings` opens the control center's Settings section: one flat list
-  of the fourteen preferences a session changes, each with its value beside the
+  of the thirteen preferences a session changes, each with its value beside the
   label, `Space`/`Enter` to change one and `Tab` for Global/Project scope. Every
   other key is a `bytefly.json` edit, listed in the configuration reference.
+- `/bytefly skills` opens the control center's Skills section: one row per Pi
+  skill with its enable checkbox, its declared description underneath, and
+  `Space` to switch it on or off for the active scope.
 - `/bytefly doctor [extension] [refresh]` prints the doctor report —
   `run_doctor`'s text — and checks that each enabled extension has what it needs
   at runtime.
 - `/bytefly update` re-runs the package update check now and reports the
   updateable packages, this package's version and release URL, or a skip.
 - `/plan <task>` starts a task in read-only planning mode.
-- `/skill:init` creates or updates `AGENTS.md` through a short built-in interview and a plan you review first.
+- `/skill:init` creates or updates the instruction file Pi reads through a short
+  built-in interview and a plan you review first.
 
 One thing the pack does on its own: once per interactive session, `bytefly`
 asks Pi's package manager which configured packages are behind, suppresses Pi's
-own startup update block, and posts one themed notice naming
-`pi update --extensions`; this package's new version also marks the header.
-Nothing about the session is sent, and `PI_OFFLINE` skips the check;
+own startup update blocks, and posts one themed notice naming
+`pi update --extensions`, plus one for a newer Pi that names `pi update` and the
+changelog; a new version also marks the header — this package's, and Pi's own.
+Nothing about the session is sent — Pi's own release request is only re-rendered —
+and `PI_OFFLINE` skips the check;
 [`docs/security.md`](./docs/security.md) covers the trust boundaries.
 
 ## What you get
 
-18 extensions, 3 packaged skills, and 2 themes, grouped by what you are trying
+19 extensions, 5 packaged skills, and 2 themes, grouped by what you are trying
 to do. Each extension's own `README.md` has the full tool and command surface.
 
 ### Plan and delegate
 
 | Extension | Tools and commands | What it does |
 |---|---|---|
-| [`plan`](./extensions/plan/) | `enter_plan_mode`, `write_plan`, `exit_plan_mode`, `set_plan_steps`, `/plan`, `--plan` | Read-only planning: the model investigates, saves a plan under `.pi/plans`, and you approve it before anything is written. Review in the TUI, or open it in a browser to leave inline comments. The `init` skill uses the same review to create or update a project's `AGENTS.md`. |
+| [`plan`](./extensions/plan/) | `enter_plan_mode`, `write_plan`, `exit_plan_mode`, `cancel_plan_mode`, `set_plan_steps`, `/plan`, `--plan` | Read-only planning: the model investigates, saves a plan under `~/.pi/agent/plans/<repo-key>` — one store per repository, shared by its worktrees — and you approve it before anything is written; with `plan.split` on, the approval can also split it into slices that separate sessions join by name. Review in the TUI, or open it in a browser to leave inline comments: the page follows the plan file live, and you can hand one reviewer link per teammate so each comments under their own name. The `init` skill uses the same review to create or update a project's `AGENTS.md`. |
 | [`check`](./extensions/check/) | `run_check`, `/check` | Detect the project's test, lint, or typecheck command (or take `check.command`), run it in a supervised process, and turn failures into a navigable `file:line` list. |
 | [`subagent`](./extensions/subagent/) | `subagent`, `/agents` | Delegate a task to a specialist agent (`explorer`, `planner`, `reviewer`, `worker`, `researcher`, `tester`, `debugger`, `documenter`) in its own Pi process — single, parallel, chained, or in the background, with a footer chip while it runs and a finish notice. |
 | [`todo`](./extensions/todo/) | `todo`, `/todos` | A persistent task list (`pending` / `in_progress` / `completed`) with a one-line widget. |
@@ -109,13 +116,14 @@ to do. Each extension's own `README.md` has the full tool and command surface.
 | [`question`](./extensions/question/) | `ask_user_question` | Ask you one or more structured questions (labelled options plus free-form "Other") and wait. |
 | [`job`](./extensions/job/) | `job`, `/jobs` | Run long-lived shell commands in the background; inspect logs, wait, or kill them. |
 | [`watch`](./extensions/watch/) | `watch`, `/watch` | Tail a file or command output, match lines against patterns, and alert you or hand matching lines to the agent. |
+| [`session-bus`](./extensions/session-bus/) | `session`, `/sessions` (screen) | Coordinate parallel Pi sessions on one repository: discover peers, exchange short messages/events, consume an inbox, join a slice of an approved split or claim paths before editing, with `/sessions` as a live board of peers, unread mail, and claims. Local, pull-based by default, with an opt-in inbox watch that raises this session's turn once it is idle; it never spawns an agent. |
 | [`context`](./extensions/context/) | `/context` | Show where the context window goes — system prompt, tool definitions, conversation by role, injected notes — and compact from the report with the before and after numbers. |
 
 ### Keep your work safe
 
 | Extension | Tools and commands | What it does |
 |---|---|---|
-| [`worktree`](./extensions/worktree/) | `enter_worktree`, `exit_worktree`, `prune_worktrees`, `list_worktrees`, `/worktree status|enter|exit|prune`, `--worktree` | Work inside an isolated `git worktree` on its own branch; the built-in read/write/edit/bash tools follow, and cleanup refuses to throw work away. |
+| [`worktree`](./extensions/worktree/) | `enter_worktree`, `exit_worktree`, `prune_worktrees`, `list_worktrees`, `/worktree` (screen), `/worktree enter|exit|prune`, `--worktree` | Work inside an isolated `git worktree` on its own branch; the built-in read/write/edit/bash tools follow, and cleanup refuses to throw work away. |
 | [`rewind`](./extensions/rewind/) | `/rewind`, `/review`, `↺ N` status chip | Snapshot the repo on every prompt under `refs/pi/rewind`, restore code, conversation, or both, and review the latest turn's changes before reverting a file. |
 | [`memory`](./extensions/memory/) | `memory`, `/memory` | Durable cross-session notes as human-editable markdown, injected into each run and managed through the `/memory` picker. |
 
@@ -124,7 +132,7 @@ to do. Each extension's own `README.md` has the full tool and command surface.
 | Extension | Tools and commands | What it does |
 |---|---|---|
 | [`web-access`](./extensions/web-access/) | `web_search`, `web_fetch` | Search the web (keyless DuckDuckGo by default, optional SearXNG or Brave) and fetch a page as readable text, with paging, `find`, per-hop SSRF checks, and optional PDF support. |
-| [`file-browse`](./extensions/file-browse/) | `/browse` | Start a loopback, read-only file browser rooted at the working directory, with Markdown preview, a fuzzy quick-open palette, and live status for a page left open. |
+| [`file-browse`](./extensions/file-browse/) | `/browse` | Start a loopback, read-only file browser rooted at the working directory, with Markdown preview, a fuzzy quick-open palette, a `path:12-30` reference copied from a selection, live status for a page left open, and a reload hint when the folder it lists changes. |
 | [`file-read`](./extensions/file-read/) | `read_doc` | Extract text from a local PDF, DOCX, DOC, ODT, RTF, or XLSX in process, paged. |
 | [`file-open`](./extensions/file-open/) | `/open`, `open_path` | Open or reveal a file or directory in the OS-native application. |
 | [`wechat`](./extensions/wechat/) | `/wechat`, `send_wechat` | Bridge a WeChat chat to the live session over the Weixin iLink bot API; inbound text, images, and files become turns. |
@@ -133,7 +141,7 @@ to do. Each extension's own `README.md` has the full tool and command surface.
 
 | Extension | Tools and commands | What it does |
 |---|---|---|
-| [`bytefly`](./extensions/bytefly/) | `/bytefly` (with the `doctor`, `settings`, and `update` subcommands), `run_doctor` | The visible layer: a themed header with a width-adaptive footer, a live run line on Pi's working row (tool, tool count, and elapsed time), and a transcript where every tool block leads with its call line and rails its result under it. `/bytefly` switches extensions and holds the curated settings; opt-in `notify` raises OS notifications when a run, a job, or a question needs you. `bytefly.json` is the config document every extension reads. |
+| [`bytefly`](./extensions/bytefly/) | `/bytefly` (with the `doctor`, `settings`, `skills`, and `update` subcommands), `run_doctor` | The visible layer: a themed header with a width-adaptive footer, a live run line on Pi's working row (tool, tool count, and elapsed time), and a transcript where every tool block leads with its call line and rails its result under it. `/bytefly` switches extensions and skills and holds the curated settings; opt-in `notify` raises OS notifications when a run, a job, or a question needs you, and Pi's own startup update blocks become themed notices and header markers. `bytefly.json` is the config document every extension reads. |
 
 ### Skills
 
@@ -143,9 +151,11 @@ advertised to the model, and its instructions load only when the task matches
 
 | Skill | What it does |
 |---|---|
-| [`init`](./skills/init/SKILL.md) | Creates or updates a project's `AGENTS.md` through a built-in interactive interview and plan review, from repository evidence. |
+| [`init`](./skills/init/SKILL.md) | Creates or updates the instruction file Pi actually reads (`AGENTS.md`, or `AGENTS.override.md` when it shadows it) through a built-in interactive interview and plan review, deriving commands, generated paths, packages, and purpose from repository evidence. |
 | [`refactor`](./skills/refactor/SKILL.md) | Audits a project or path for refactoring opportunities: duplication, oversized modules, coupling, drift, and high-churn code with weak tests. Returns a ranked shortlist with evidence and verification commands; discovery is read-only. |
-| [`changelog`](./skills/changelog/SKILL.md) | Drafts the candidate bullet before implementation, records one user-visible bullet in the same commit, audits commits since the last tag for missing or weak entries, and curates a release section from the project's own rules. |
+| [`changelog`](./skills/changelog/SKILL.md) | Records one user-visible bullet in the same commit as the change, audits commits since the last tag for missing or weak entries, and curates a release section from the project's own rules. |
+| [`parallel-dev`](./skills/parallel-dev/SKILL.md) | Runs parallel development as a lead and a team: a lead accountable for the contract, shared files, merge order and a green baseline who also owns a workstream, one owner per disjoint workstream, a walking skeleton first, advisory path claims, a turn-boundary protocol, and independent verification. |
+| [`next`](./skills/next/SKILL.md) | Suggests the next todos by priority — loose work, a broken or unfinished promise — from this conversation once it has done work, and from the tree only when it is empty, with the words and a path after the command saying what to ask and where to look. Read-only, and capped at a short numbered list with evidence and the command that starts each. |
 
 ### Themes
 
@@ -168,6 +178,10 @@ README use `bytefly-dark`.
 <td width="50%"><img src="assets/job.png" alt="The /jobs screen: a summary line, the job rows with status glyphs and elapsed times, and the focused job's command, cwd, and last output line" width="100%"></td>
 <td width="50%"><img src="assets/rewind.png" alt="The /rewind screen: prompt rows newest first, a diamond marking prompts with a code snapshot, and the focused prompt's snapshot detail" width="100%"></td>
 </tr>
+<tr>
+<td width="50%"><img src="assets/session.png" alt="The /sessions screen: a counts summary, session rows with unread and claim counts, the focused session's cwd, claims, and identity, and the Tab scope and watch hints" width="100%"></td>
+<td width="50%"><img src="assets/watch.png" alt="The /watch screen: a summary line and watcher rows with live match counts and the focused watcher's pattern and source" width="100%"></td>
+</tr>
 </table>
 
 ### Tool renderings
@@ -178,7 +192,7 @@ each tool's own renderer (or bytefly's generic fallback): the user nameplate and
 a notice, Pi's built-ins, and every tool this package registers. The rail glyph
 is the `chrome.rail` setting.
 
-![A full-page reference of every tool block: the user nameplate, a notice, the turn rule, Pi's read, write, edit, bash, grep, find, and ls renderers, and every package tool — todo, goal, memory, job, watch, subagent, web_search, web_fetch, read_doc, open_path, run_doctor, run_check, send_wechat, ask_user_question, the plan tools, and the worktree tools](assets/tool-renderings.png)
+![A full-page reference of every tool block: the user nameplate, a notice, the turn rule, Pi's read, write, edit, bash, grep, find, and ls renderers, and every package tool — todo, goal, memory, job, watch, subagent, web_search, web_fetch, read_doc, open_path, run_doctor, run_check, send_wechat, ask_user_question, session, the plan tools, and the worktree tools](assets/tool-renderings.png)
 
 `bun run screenshot:tools` regenerates it; `bun run screenshot:tool-pngs` writes
 one image per tool under [`assets/tools/`](./assets/tools).

package/README.zh-CN.md

@@ -33,14 +33,15 @@ Pi 提供一个精简而锐利的内核。这个包是在其之上的一层，
 - **先规划，再执行。** 只读的规划模式把一次任务变成一份可审阅的计划文件，智能体
   随后按步骤执行。
 - **委派与并行。** 在各自的 Pi 进程中运行专职子智能体，把耗时构建交给后台任务，
-  让 watcher 盯住日志直到出现关键行 —— 你继续把握主线。
+  让 watcher 盯住日志直到出现关键行，还能让同一仓库上的多个会话通过共享总线协作、
+  各自负责自己的一摊 —— 你继续把握主线。
 - **绝不搞乱检出区。** 隔离只需一条命令：任务切到自己分支的 `git worktree` 中，
   每次提示还会生成快照，因此一次糟糕的回合只需一个 `/rewind` 就能回退。
 - **跨会话记忆。** 持久笔记以纯 Markdown 保存，可读、可改、可纳入版本控制。
 - **连通网络与文件。** 在会话内搜索、抓取、浏览、打开和读取文档与表格，还可通过桥接
   在微信里直接回答当前会话。
 
-一切都是可选的，且可独立开关。没有任何东西被分叉，也没有修改已安装的 Pi 文件：
+一切都是可选的，且可独立开关。没有任何东西被分叉，也没有写入任何已安装的 Pi 文件：
 Pi 通过 [`package.json`](./package.json) 里的 `pi` manifest 发现这些资源，因此
 该包能像其他包一样随 Pi 一起升级。
 
@@ -64,31 +65,35 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 
 - `/bytefly [extension] [refresh]` 打开控制中心的 Extensions 区，列出每个已加载扩展
   及其运行时健康状况，启用或禁用一个扩展，并用 `x` 清除项目级覆盖。
-- `/bytefly settings` 打开控制中心的 Settings 区：一次会话会改动的十四项设置排成
+- `/bytefly settings` 打开控制中心的 Settings 区：一次会话会改动的十三项设置排成
   一列，值紧跟在名称之后；用 `Space`/`Enter` 修改，用 `Tab` 切换 Global/Project
   作用域。其余键都是 `bytefly.json` 编辑，见配置参考。
+- `/bytefly skills` 打开控制中心的 Skills 区：每个 Pi 技能一行，带启用复选框与
+  `SKILL.md` 声明的描述，用 `Space` 在当前作用域开关它。
 - `/bytefly doctor [extension] [refresh]` 打印 doctor 报告（`run_doctor` 的文本），
   检查每个已启用扩展在运行时是否具备所需条件。
 - `/bytefly update` 立即重跑包更新检查，报告可更新的包、本包版本与 release URL，
   或跳过原因。
 - `/plan <任务>` 以只读规划模式开始一个任务。
-- `/skill:init` 通过内置的简短访谈和先审阅的计划创建或更新 `AGENTS.md`。
+- `/skill:init` 通过内置的简短访谈和先审阅的计划创建或更新 Pi 实际读取的指令文件。
 
 还有一件包会自行做的事：每次交互式会话启动时，`bytefly` 询问 Pi 的包管理器
-哪些已配置的包有更新，抑制 Pi 自己的启动更新块，并用一个主题化 notice 列出它们
-和 `pi update --extensions`；本包的新版本还会标在头部。不会发送任何会话内容，
-`PI_OFFLINE` 可跳过检查；信任边界见 [`docs/security.md`](./docs/security.md)。
+哪些已配置的包有更新，抑制 Pi 自己的启动更新块，并用主题化 notice 列出它们
+和 `pi update --extensions`；Pi 自身的新版本也由它绘制，提示 `pi update` 与
+changelog。新版本还会标在头部 —— 本包的与 Pi 自己的。不会发送任何会话内容
+（Pi 自己的版本请求只是被重新绘制），`PI_OFFLINE` 可跳过检查；信任边界见
+[`docs/security.md`](./docs/security.md)。
 
 ## 包含内容
 
-18 个扩展、3 个随包技能，以及 2 个主题，按你想达成的目标分组。每个扩展自己的 `README.md`
+19 个扩展、5 个随包技能，以及 2 个主题，按你想达成的目标分组。每个扩展自己的 `README.md`
 提供完整的工具与命令说明。
 
 ### 规划与委派
 
 | 扩展 | 工具与命令 | 说明 |
 |---|---|---|
-| [`plan`](./extensions/plan/) | `enter_plan_mode`, `write_plan`, `exit_plan_mode`, `set_plan_steps`, `/plan`, `--plan` | 只读规划：模型先调研，把计划保存到 `.pi/plans` 下，你在写入任何内容前先批准。可在 TUI 中审阅，或在浏览器中打开以留下行内评论。`init` 技能复用同一审阅流程来创建或更新项目的 `AGENTS.md`。 |
+| [`plan`](./extensions/plan/) | `enter_plan_mode`, `write_plan`, `exit_plan_mode`, `cancel_plan_mode`, `set_plan_steps`, `/plan`, `--plan` | 只读规划：模型先调研，把计划保存到 `~/.pi/agent/plans/<repo-key>` 下（每个仓库一份，各 worktree 共享），你在写入任何内容前先批准；开启 `plan.split` 后，批准时还可以把计划拆成切片，由独立会话按名字加入。可在 TUI 中审阅，或在浏览器中打开以留下行内评论：页面会跟随计划文件实时更新，你也可以给每位同事各一个审阅链接，各自以绑定的名义评论。`init` 技能复用同一审阅流程来创建或更新项目的 `AGENTS.md`。 |
 | [`check`](./extensions/check/) | `run_check`, `/check` | 检测项目的测试、lint 或类型检查命令（或使用 `check.command`），在受监管的进程中运行，并把失败整理成可导航的 `file:line` 列表。 |
 | [`subagent`](./extensions/subagent/) | `subagent`, `/agents` | 把任务委派给专职智能体（`explorer`、`planner`、`reviewer`、`worker`、`researcher`、`tester`、`debugger`、`documenter`），运行在它自己的 Pi 进程中 —— 单个、并行、链式或在后台运行；后台运行时在页脚显示数量，完成后发出提示。 |
 | [`todo`](./extensions/todo/) | `todo`, `/todos` | 持久任务列表（`pending` / `in_progress` / `completed`），配一行式 widget。 |
@@ -96,13 +101,14 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 | [`question`](./extensions/question/) | `ask_user_question` | 向你提出一个或多个结构化问题（带标签的选项外加自由填写的“其他”）并等待回答。 |
 | [`job`](./extensions/job/) | `job`, `/jobs` | 在后台运行长时 shell 命令；查看日志、等待或结束它们。 |
 | [`watch`](./extensions/watch/) | `watch`, `/watch` | 跟踪文件或命令输出，按模式匹配行，并提醒你或把匹配的行交给智能体。 |
+| [`session-bus`](./extensions/session-bus/) | `session`, `/sessions`（屏幕） | 在同一仓库的多个 Pi 会话之间协调：发现彼此、交换简短消息/事件、读取收件箱、加入已批准拆分中的一个切片或编辑前声明路径；`/sessions` 是显示对等会话、未读邮件和路径声明的实时面板。本地、默认拉取式；会话可以自行开启收件箱监视，在空闲时唤起本会话的一轮，从不创建智能体。 |
 | [`context`](./extensions/context/) | `/context` | 显示上下文窗口的去向 —— 系统提示、工具定义、按角色拆分的对话、注入的笔记 —— 并从报告中触发压缩，给出压缩前后的数字。 |
 
 ### 保护你的工作
 
 | 扩展 | 工具与命令 | 说明 |
 |---|---|---|
-| [`worktree`](./extensions/worktree/) | `enter_worktree`, `exit_worktree`, `prune_worktrees`, `list_worktrees`, `/worktree status|enter|exit|prune`, `--worktree` | 在隔离的 `git worktree` 里、在独立分支上工作；内置的 read/write/edit/bash 工具会跟随，清理时也不会丢弃工作。 |
+| [`worktree`](./extensions/worktree/) | `enter_worktree`, `exit_worktree`, `prune_worktrees`, `list_worktrees`, `/worktree` (screen), `/worktree enter|exit|prune`, `--worktree` | 在隔离的 `git worktree` 里、在独立分支上工作；内置的 read/write/edit/bash 工具会跟随，清理时也不会丢弃工作。 |
 | [`rewind`](./extensions/rewind/) | `/rewind`, `/review`, `↺ N` 状态标记 | 每次提示时在 `refs/pi/rewind` 下对仓库生成快照，可恢复代码、对话或两者，并在还原文件前审阅本轮改动。 |
 | [`memory`](./extensions/memory/) | `memory`, `/memory` | 跨会话的持久笔记，以可人工编辑的 Markdown 保存，注入每次运行，并通过 `/memory` 选择器管理。 |
 
@@ -111,7 +117,7 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 | 扩展 | 工具与命令 | 说明 |
 |---|---|---|
 | [`web-access`](./extensions/web-access/) | `web_search`, `web_fetch` | 搜索网络（默认无密钥的 DuckDuckGo，可选 SearXNG 或 Brave），并把页面抓取为可读文本，支持分页、`find`、逐跳 SSRF 检查以及可选的 PDF 支持。 |
-| [`file-browse`](./extensions/file-browse/) | `/browse` | 启动一个以工作目录为根的本地回环、只读文件浏览器，带 Markdown 预览、模糊快速打开面板，以及为常开页面提供的实时状态。 |
+| [`file-browse`](./extensions/file-browse/) | `/browse` | 启动一个以工作目录为根的本地回环、只读文件浏览器，带 Markdown 预览、模糊快速打开面板、从选区复制 `path:12-30` 引用、为常开页面提供的实时状态，以及所列目录变化时的重新加载提示。 |
 | [`file-read`](./extensions/file-read/) | `read_doc` | 在进程内从本地 PDF、DOCX、DOC、ODT、RTF 或 XLSX 提取文本，支持分页。 |
 | [`file-open`](./extensions/file-open/) | `/open`, `open_path` | 用操作系统原生应用打开或定位文件/目录。 |
 | [`wechat`](./extensions/wechat/) | `/wechat`, `send_wechat` | 通过微信 iLink 机器人 API 把微信聊天桥接到实时会话；入站文本、图片和文件会变成回合。 |
@@ -120,7 +126,7 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 
 | 扩展 | 工具与命令 | 说明 |
 |---|---|---|
-| [`bytefly`](./extensions/bytefly/) | `/bytefly`（`doctor`、`settings`、`update` 子命令）、`run_doctor` | 可见的那一层：主题化头部配宽度自适应页脚，Pi 工作行上实时显示工具、调用次数与计时，转录中每个工具块以调用行为标题、结果正文挂在其下方。`/bytefly` 可开关扩展并集中存放精选设置；可选的 `notify` 会在一次运行、后台任务或待回答的问题需要你时发出系统通知。`bytefly.json` 是每个扩展都会读取的配置文档。 |
+| [`bytefly`](./extensions/bytefly/) | `/bytefly`（`doctor`、`settings`、`skills`、`update` 子命令）、`run_doctor` | 可见的那一层：主题化头部配宽度自适应页脚，Pi 工作行上实时显示工具、调用次数与计时，转录中每个工具块以调用行为标题、结果正文挂在其下方。`/bytefly` 可开关扩展与技能，并集中存放精选设置；可选的 `notify` 会在一次运行、后台任务或待回答的问题需要你时发出系统通知，Pi 自己的启动更新块也会变成主题化 notice 与头部标记。`bytefly.json` 是每个扩展都会读取的配置文档。 |
 
 ### 技能
 
@@ -128,9 +134,11 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 
 | 技能 | 说明 |
 |---|---|
-| [`init`](./skills/init/SKILL.md) | 基于仓库证据，通过内置的交互式访谈和计划审阅创建或更新项目的 `AGENTS.md`。 |
+| [`init`](./skills/init/SKILL.md) | 通过内置的交互式访谈和计划审阅，基于仓库证据创建或更新 Pi 实际读取的指令文件（`AGENTS.md`；存在覆盖它的 `AGENTS.override.md` 时则是该文件），并从仓库推导命令、生成路径与包范围。 |
 | [`refactor`](./skills/refactor/SKILL.md) | 审计项目或路径中的重构机会：重复代码、过大的模块、耦合、结构漂移，以及测试薄弱且改动频繁的代码。返回带证据和验证命令的排序清单；发现阶段只读。 |
-| [`changelog`](./skills/changelog/SKILL.md) | 实现前先起草候选条目，实施后在同一次提交中记录一条用户可见的变更，审计上一个 tag 以来缺失或薄弱的条目，并按项目自己的规则整理发布段落。 |
+| [`changelog`](./skills/changelog/SKILL.md) | 在变更的同一次提交中记录一条用户可见的条目，审计上一个 tag 以来缺失或薄弱的条目，并按项目自己的规则整理发布段落。 |
+| [`parallel-dev`](./skills/parallel-dev/SKILL.md) | 以“负责人加团队”的方式组织并行开发：负责人对契约、共享文件、合并顺序和绿色基线负责，同时自己承担一个工作流；每个互不重叠的工作流各有一名所有者，先做可运行的纵向骨架，声明路径，在每个回合的边界协作，并由他人独立验证。 |
+| [`next`](./skills/next/SKILL.md) | 按优先级建议接下来的待办——未收尾的工作、已破坏或未兑现的承诺——会话已有工作时基于本次会话，仅当会话为空时才基于代码树；命令后的文字说明问什么、路径说明看哪里。只读，输出带证据和启动命令的简短编号清单。 |
 
 ### 主题
 
@@ -152,6 +160,10 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 <td width="50%"><img src="assets/job.png" alt="/jobs 屏幕：摘要行、带状态字形和耗时的任务行，以及焦点任务的命令、工作目录和最后一行输出" width="100%"></td>
 <td width="50%"><img src="assets/rewind.png" alt="/rewind 屏幕：从新到旧的提示行、标记有代码快照的菱形，以及焦点提示的快照详情" width="100%"></td>
 </tr>
+<tr>
+<td width="50%"><img src="assets/session.png" alt="/sessions 屏幕：计数摘要、带未读数和路径声明数的会话行、焦点会话的工作目录、声明和身份，以及 Tab 范围与监视提示" width="100%"></td>
+<td width="50%"><img src="assets/watch.png" alt="/watch 屏幕：摘要行和带实时匹配数的监视行，以及焦点监视的模式与来源" width="100%"></td>
+</tr>
 </table>
 
 ### 工具渲染
@@ -161,7 +173,7 @@ Windows、macOS 和 Linux 均可运行，完整检查套件在每个平台上都
 （每个工具自身的渲染器或 bytefly 的通用回退）——用户铭牌与通知、Pi 内置工具，以及本包注册的每个工具。
 侧栏字形由 `chrome.rail` 设置。
 
-![每个工具块的全页参考图：用户铭牌、通知、回合分隔线，Pi 的 read、write、edit、bash、grep、find、ls 渲染器，以及每个包内工具——todo、goal、memory、job、watch、subagent、web_search、web_fetch、read_doc、open_path、run_doctor、send_wechat、ask_user_question、plan 工具与 worktree 工具](assets/tool-renderings.png)
+![每个工具块的全页参考图：用户铭牌、通知、回合分隔线，Pi 的 read、write、edit、bash、grep、find、ls 渲染器，以及每个包内工具——todo、goal、memory、job、watch、subagent、web_search、web_fetch、read_doc、open_path、run_doctor、send_wechat、ask_user_question、session、plan 工具与 worktree 工具](assets/tool-renderings.png)
 
 `bun run screenshot:tools` 重新生成它；`bun run screenshot:tool-pngs` 在
 [`assets/tools/`](./assets/tools) 下为每个工具输出一张图。