agent-from-zero 0.27.0 → 0.29.0

package/README.md

@@ -45,48 +45,53 @@ MINI_AGENT_MODEL=gpt-4.1-mini
 
 ## 进度
 
-| Tag | 内容 |
+| Tag | What it adds |
 |---|---|
-| `day1` | ~50 行 raw tool-calling loop:一个工具、一个循环、两条保险 |
-| `day2` | 工具五件套(read/write/edit/search/bash):编辑前必须先读、old_string 唯一性、永不抛异常的 dispatch、四要素工具说明书 |
-| `day3` | 主循环状态机:错误分类、指数退避(带抖动)、双重重试预算、熔断器,8 种终态各配人话提示,Ctrl+C 体面退出 |
-| `day4` | 权限系统(allow/ask/deny):禁飞区硬拦(.git/.ssh/.env)、deny 永远赢(bypass 也压不过)、非交互 fail closed、bash 按输入定危险等级、秘密文件不进上下文 |
-| `day5` | 上下文管理:token 估算(字节法,宁多勿少)、80% 触发线、6 段结构化摘要(双保险禁工具)、文件恢复(最近 5 个,从磁盘重读)、主动+被动压缩、压缩熔断 |
-| `day6` | 多轮 REPL(行队列输入、/clear、Ctrl+C 三态语义)、system prompt 宪法(数字锚/反面禁令/防虚报)、AGENT.md 项目记忆、流式输出 + 双重看门狗(90s 断流/30s 只警告) |
-| `day7` | 子 agent(task 工具:独立上下文、一层分身不套娃、报告是材料不是结论、文件状态快照隔离)+ 最小 eval(10 道固定题、物证判分、8/10 及格线,`npm run eval`) |
-| `day8` | 封装发布:`bin` 入口 + tsc 构建,npx 一行启动(包名 `agent-from-zero`);内置 .env 加载;缺 key 人话报错 |
-| `day9` | 教程本身:11 章中文教程开源进 [docs/](docs/README.md) |
-| `day10` | Bring your own model:三个环境变量换任意 OpenAI 兼容端点,`npm run eval` 兼职模型横评 |
-| `day11` | 工业级 CLI 表面:settings.json 双层配置(用户权限规则,deny 永远赢)、slash 命令(/help /model /compact)、`-p` 单发模式、`--version`、测试套件进仓库(`npm test`,60 例) |
-| `day12` | 会话持久化(`--resume`,快照式 + 原子写)+ 本地遥测(事件 JSONL + `/stats`,绝不出本机,`MINI_AGENT_NO_TELEMETRY=1` 关闭) |
-| `day13` | 异步执行:run_bash 换 spawn(不冻 UI、Ctrl+C 秒杀、30s 超时 SIGKILL)、超长输出落盘、只读工具并行;`npm test` 65 例 |
-| `day14` | Hook 生命周期系统:PreToolUse(拦截工具)/PostToolUse(追加校验)/SessionStart(注入上下文)/Stop(测试驱动 AI),退出码协议、stderr 喂回模型;`npm test` 76 例 |
-| `day15` | MCP 接入:stdio JSON-RPC 客户端(握手/行缓冲/id 配对),外部 server 工具注册成内置同款(同一 dispatch + 权限闸门,默认 ask);`npm test` 84 例 |
-| `day16` | 可选 LLM 权限裁判:只看规则判 ask 的灰色地带、只能 ask→allow、prompt 铁死格式、解析 fail closed、自带熔断、默认关闭;`npm test` 98 例 |
-| `day17` | 长期记忆:跨会话的 MEMORY.md(人类可读)+ remember 工具,作为「背景」注入(区别于 AGENT.md 的「命令」),有界 30 条/8KB 且自修复;`npm test` 114 例 |
-| `day18` | 成本核算:从流里捕获 usage(include_usage)、抹平 DeepSeek/OpenAI 格式差异、纯函数算成本与缓存命中率、`/cost` 看花销(命中率即缓存击穿预警);`npm test` 130 例 |
-| `day19` | 工具调用校验:够用就好的 schema 校验(必填/类型/形状)、钉在 dispatch 唯一入口、批量返回精确修复指引、自动覆盖内置+MCP 工具(送出前拦下坏参数);`npm test` 146 例 |
-| `day20` | 计划模式:只读研究、想清楚再交计划由你批准;权限层一个外层过滤器(只收紧不放松、盖过预批准、保留禁飞区理由)+ `exit_plan_mode` 工具借用 ask 流程闭合审批 + `/plan` 开关;`npm test` 167 例 |
-| `day21` | diff 预览:写文件前把真实改动(绿 +/红 -)摊给你看再批准;十几行 LCS 行级 diff、折叠未变动行、复现工具语义无副作用推演、永不抛异常 fail 到 null;在闸门后执行前打印;`npm test` 189 例 |
-| `day22` | `/undo` 撤销上一次写入:写入前埋一行存「之前」快照(有界栈),`/undo` 还原内容或删掉新建文件、复用 Day 21 diff 画出撤销结果;`/clear` 同时清栈;`npm test` 204 例 |
-| `day23` | 子 agent 模型分级:`subAgentModel` 配置让委托跑在不同档位(便宜模型干杂活 / 强模型当顾问审查),纯函数 `subAgentModelFor` 定回落规则,首次用假 client 端到端验证「编排→委托→回归」真换了模型;`npm test` 212 例 |
-| `day24` | 远程 MCP(HTTP/SSE):把 Day 15 的传输从协议里剥成 Transport 接口,新增 HTTP 传输(POST→按 Content-Type 分 JSON/SSE 解析、回带 Mcp-Session-Id、AbortSignal 超时),配置 `url` 即走远程;协议层/发现/注册零改动;`npm test` 218 例 |
-| `day25` | `/diff` 会话改动总览:撤销账本旁加一张「会话起点基线」表(每个文件首次被碰的样子,只记一次),`/diff` 与磁盘现状逐个比、复用 Day 21 渲染净改动、净零跳过,三态 created/modified/deleted;`npm test` 227 例 |
-| `day26` | `@file` 引用:提问里 `@路径` 即把文件内容附进消息省一次 read_file 往返;认行首/空白后的 `@`、剥句尾标点;每个提及过 read_file 权限闸门——密钥文件(.env/*.pem)拒绝不读、被拒出声、没命中静默;`npm test` 244 例 |
-| `day27` | `/resume` 会话选择器:`session.ts` 加 listSessions(摘要)/loadSession(按 id)/sessionTitle(首句当标题、剥 @file 附件),REPL 列表+选号+整体切换(重建宪法、换 id、清读态与撤销栈),会话目录改惰性求值以可测;`npm test` 256 例 |
-| `day28` | 终端 UI:新增 `ui.ts` 收拢外观——圆角欢迎框、`❯` 提示符、`⏺`/`⎿` 活动标记(替 emoji)、走秒 spinner(搭车看门狗);讲透「颜色码污染宽度计算」要先按纯文本对齐再上色;`npm test` 277 例 |
-| `day29` | 方向键审批菜单:审批从「打 y/N」换成 ↑↓ 选(`menu.ts` 借 stdin 进 raw 模式、原地重绘、选完归还),带编号+底部操作提示;中间项「本会话别再问」复用 allow 规则(run_bash 只升 ask 不碰硬 deny);非 TTY 回退打字;`npm test` 289 例 |
-| `day30` | 多问题选择表单:给模型一个 `ask_user` 工具,需用户决策时弹表单(`form.ts` 纯状态机:扁平光标+reduce+提交校验+缺口跳转,`promptForm` 薄键盘层),收集成 question→answer 喂回;输入区改成 Claude Code 式「分隔线+❯」;`npm test` 309 例 |
-| `day31` | 运行时切换模型:`/model` 从只读升级成可切——`/model <name>` 直接设,裸 `/model` 拉端点 `/models` 列表用 Day 29 菜单选;切换只改 `CONFIG.model`、靠主循环每轮重读立即生效;`npm test` 314 例 |
-| `day32` | 让换模型看得见:流式循环多读 `reasoning_content`,把推理模型(R1)的思考灰着流出来(只渲染不进历史,绝不回传);spinner 显示当前模型作为切换生效的证明;`npm test` 324 例 |
-| `day33` | Skill 系统(Markdown 即插件):`SKILL.md`(frontmatter+正文)=一个插件、执行者是模型;渐进披露(列表进上下文/正文按需加载)、`skill` 工具说明书即列表、`disableModelInvocation` 分模型可调/仅用户、`/skills` `/skill <name>`;`npm test` 360 例 |
-| `day34` | 记忆深化:四类型(user/feedback/project/reference,feedback 最值钱)、每轮自动提取(便宜模型读对话存「用户纠正过的」,默认关、永不抛异常)、按类型分组注入、向后兼容旧格式;`npm test` 375 例 |
-| `day35` | Hook 深化:4→10 个时机(加 UserPromptSubmit/SessionEnd/Pre·PostCompact/Subagent*),配置改开放 record、exit 2 语义按事件由调用方解读、SessionEnd 极短超时;**`PreToolUse` 可改写工具参数**(权限三态做不到的「岔道」,严格在安全检查之后);`npm test` 381 例 |
+| `day1` | ~50-line raw tool-calling loop: one tool, one loop, two safety nets (errors fed back, a round cap) |
+| `day2` | The five core tools (read/write/edit/search/bash): read-before-edit, `old_string` uniqueness, a `dispatch` that never throws, the four-part tool manual |
+| `day3` | Main loop as a state machine: error classification, exponential backoff (with jitter), dual retry budgets, a circuit breaker, 8 terminal states each with a plain-English note, graceful Ctrl+C exit |
+| `day4` | Permission system (allow/ask/deny): hard-blocked no-fly zones (.git/.ssh/.env), deny always wins (even bypass can't override), fail-closed when non-interactive, bash risk graded by the command, secret files kept out of context |
+| `day5` | Context management: token estimation (byte-based, err high), an 80% trigger line, a 6-section structured summary (tools double-disabled), file recovery (last 5, re-read from disk), proactive + reactive compaction, a compaction circuit breaker |
+| `day6` | Multi-turn REPL (line-queue input, `/clear`, three-state Ctrl+C), a system-prompt constitution (number anchors / negative prohibitions / no false claims), AGENT.md project memory, streaming output + a dual watchdog (90s stall / 30s warn-only) |
+| `day7` | Sub-agents (the `task` tool: isolated context, one level deep with no nesting, the report is material not the verdict, file-state snapshot isolation) + a minimal eval (10 fixed tasks, evidence-based scoring, 8/10 pass bar, `npm run eval`) |
+| `day8` | Packaging & release: a `bin` entry + tsc build, one-line `npx` launch (package `agent-from-zero`); built-in .env loading; a plain-English error when the key is missing |
+| `day9` | The tutorial itself: 11 Chinese chapters open-sourced into [docs/](docs/README.md) |
+| `day10` | Bring your own model: three env vars swap in any OpenAI-compatible endpoint; `npm run eval` doubles as a model bake-off |
+| `day11` | Industrial-grade CLI surface: two-layer settings.json (user permission rules, deny always wins), slash commands (/help /model /compact), `-p` one-shot mode, `--version`, a test suite in the repo (`npm test`, 60 cases) |
+| `day12` | Session persistence (`--resume`, snapshot-style + atomic writes) + local telemetry (event JSONL + `/stats`, never leaves the machine, off via `MINI_AGENT_NO_TELEMETRY=1`) |
+| `day13` | Async execution: run_bash switched to spawn (no UI freeze, instant Ctrl+C, 30s timeout then SIGKILL), oversized output spilled to disk, read-only tools run in parallel; `npm test` 65 cases |
+| `day14` | Hook lifecycle system: PreToolUse (intercept a tool) / PostToolUse (append checks) / SessionStart (inject context) / Stop (test-driven AI), exit-code protocol, stderr fed back to the model; `npm test` 76 cases |
+| `day15` | MCP integration: a stdio JSON-RPC client (handshake / line buffering / id pairing), external-server tools registered just like built-ins (same dispatch + permission gate, ask by default); `npm test` 84 cases |
+| `day16` | Optional LLM permission judge: only for the ask gray-zone the rules left undecided, can only ask→allow, a rigid prompt format, fail-closed parsing, its own circuit breaker, off by default; `npm test` 98 cases |
+| `day17` | Long-term memory: a cross-session MEMORY.md (human-readable) + a remember tool, injected as "background" (vs AGENT.md's "commands"), bounded at 30 entries / 8KB and self-healing; `npm test` 114 cases |
+| `day18` | Cost accounting: capture usage from the stream (include_usage), smooth over DeepSeek/OpenAI format differences, pure-function cost + cache-hit-rate math, `/cost` to see spend (the hit rate doubles as a cache-bust warning); `npm test` 130 cases |
+| `day19` | Tool-call validation: just-enough schema checks (required / type / shape), pinned at dispatch's single entry, batched precise fix-it guidance, auto-covering built-in + MCP tools (bad args caught before they're sent); `npm test` 146 cases |
+| `day20` | Plan mode: read-only research, think it through, hand you a plan to approve; one outer filter in the permission layer (only tightens, never loosens, overrides pre-approvals, keeps no-fly reasons) + an `exit_plan_mode` tool that borrows the ask flow to close approval + a `/plan` toggle; `npm test` 167 cases |
+| `day21` | Diff preview: before a write, show the real change (green +/red -) for you to approve; a ~dozen-line LCS line-diff, collapsed unchanged lines, side-effect-free replay of tool semantics, never throws (fails to null); printed after the gate, before execution; `npm test` 189 cases |
+| `day22` | `/undo` reverts the last write: stash a "before" snapshot ahead of each write (a bounded stack), `/undo` restores the content or deletes a newly created file, reusing the Day 21 diff to draw the undo; `/clear` also clears the stack; `npm test` 204 cases |
+| `day23` | Sub-agent model tiering: a `subAgentModel` config runs delegation on a different tier (cheap model for chores / strong model as a reviewer), a pure `subAgentModelFor` defines the fallback, first end-to-end proof via a fake client that "orchestrate→delegate→regress" really switched models; `npm test` 212 cases |
+| `day24` | Remote MCP (HTTP/SSE): peel Day 15's transport out of the protocol into a Transport interface, add an HTTP transport (POST → parse JSON/SSE by Content-Type, carry Mcp-Session-Id, AbortSignal timeout); set `url` to go remote; zero changes to the protocol / discovery / registration; `npm test` 218 cases |
+| `day25` | `/diff` session change overview: alongside the undo ledger, a "session-start baseline" table (each file as first touched, recorded once), `/diff` compares each against disk, reuses Day 21 to render net changes, skips net-zero, three states created/modified/deleted; `npm test` 227 cases |
+| `day26` | `@file` mentions: an `@path` in your question attaches the file's content to the message, saving a read_file round-trip; recognizes `@` at line start / after whitespace, strips trailing punctuation; each mention passes the read_file gate — secret files (.env/*.pem) are refused, refusals voiced, misses silent; `npm test` 244 cases |
+| `day27` | `/resume` session picker: `session.ts` gains listSessions (summaries) / loadSession (by id) / sessionTitle (first line as title, strips @file attachments), a REPL list + pick-by-number + wholesale switch (rebuild the constitution, swap the id, clear read-state and the undo stack), sessions dir made lazy for testability; `npm test` 256 cases |
+| `day28` | Terminal UI: a new `ui.ts` owns the look — a rounded welcome box, the `❯` prompt, `⏺`/`⎿` activity markers (replacing emoji), a ticking spinner (riding the watchdog); explains why "color codes pollute width math" so you align on plain text first, then color; `npm test` 277 cases |
+| `day29` | Arrow-key approval menu: approval moves from "type y/N" to ↑↓ select (`menu.ts` borrows stdin into raw mode, redraws in place, hands it back), numbered + a footer hint; the middle "don't ask again this session" reuses the allow rule (run_bash only escalates ask, never touches hard deny); non-TTY falls back to typing; `npm test` 289 cases |
+| `day30` | Multi-question form: give the model an `ask_user` tool that pops a form when it needs your decision (`form.ts` pure state machine: flat cursor + reduce + submit validation + jump-to-gap, `promptForm` a thin keyboard layer), collected into question→answer fed back; the input area becomes Claude-Code-style "rule + ❯"; `npm test` 309 cases |
+| `day31` | Switch models at runtime: `/model` upgrades from read-only to switchable — `/model <name>` sets it, bare `/model` pulls the endpoint's `/models` list into the Day 29 menu; the switch only changes `CONFIG.model` and takes effect immediately via the loop re-reading it each round; `npm test` 314 cases |
+| `day32` | Make a model switch visible: the streaming loop also reads `reasoning_content`, streaming a reasoning model's (R1) thinking dimmed (render-only, never into history, never sent back); the spinner shows the current model as proof the switch took; `npm test` 324 cases |
+| `day33` | Skill system (Markdown as a plugin): a `SKILL.md` (frontmatter + body) = one plugin, executed by the model; progressive disclosure (the listing in context / the body loaded on demand), the `skill` tool manual is the listing, `disableModelInvocation` splits model-callable vs user-only, `/skills` `/skill <name>`; `npm test` 360 cases |
+| `day34` | Deeper memory: four types (user/feedback/project/reference, feedback the most valuable), auto-extraction each turn (a cheap model reads the conversation and stores "what the user corrected", off by default, never throws), grouped injection by type, backward-compatible with the old format; `npm test` 375 cases |
+| `day35` | Deeper hooks: 4→10 events (adding UserPromptSubmit/SessionEnd/Pre·PostCompact/Subagent*), config becomes an open record, exit-2 meaning interpreted per-event by the caller, an extra-short SessionEnd timeout; **`PreToolUse` can rewrite a tool's arguments** (the "side road" the three-state permission can't do, strictly after the safety check); `npm test` 381 cases |
+| `day36` | TodoWrite: a `todo_write` tool that lets the agent keep its own checklist (whole-list replace, at most one in_progress, ✓/▶/○ render, `/todos` to view) — planning, not new powers; plus a nag that injects a reminder when an unfinished plan goes stale for N rounds, so long tasks don't drift off the original goal; `npm test` 474 cases |
+| `day37` | Background tasks: a `run_bash_background` tool spawns a slow command (install/build/test/dev server) and returns a `bg_*` id immediately instead of blocking on run_bash's 30s watchdog — the agent keeps reasoning while it runs; when it finishes the loop injects a `<task_notification>` as the next turn (once per task), a `bash_output` tool polls a job's new output (essential for servers that never "finish"), `/bg` lists them and they're SIGKILLed on exit; same permission gate as run_bash, so the danger rules are identical; `npm test` 502 cases |
+| `day38` | Agent teams: past the one-shot sub-agent — a `spawn_teammate` tool starts PERSISTENT teammates that run concurrently (fire-and-forget async loops interleaving on the one event loop) and coordinate through `send_message`; mailboxes are real `.jsonl` files on disk (`.mini-agent/team/`, consumptive read = observable, no lock needed on a single event loop), the loop injects each agent's inbox as a turn and the lead WAITS on the team instead of ending early, teammates are bounded (round cap) with a focused toolset and a non-interactive permission policy (writes auto-proceed, risky bash declined, deny still wins), `/team` lists them; `npm test` 523 cases |
+| `day39` | Team protocols: explicit message contracts, not vibes — every protocol is the same request→response correlated by a `request_id`, tracked as `ProtocolState` (pending→approved/rejected) with a `match_response` that rejects a mismatched reply (a shutdown_response can't approve a plan). Two contracts: **graceful shutdown** (lead `request_shutdown` → teammate finishes, confirms, exits — no orphans/half-writes) and **plan approval** (teammate `submit_plan` → lead `review_plan` before risky work runs). Teammates go from bounded auto-exit to an IDLE LOOP (wait for work, leave on the shutdown handshake; bounded backstop so an abandoned one self-exits), and the lead disbands the team gracefully when it finishes; the inbox-consume step routes protocol messages then hands the rest to the model; `npm test` 545 cases |
+| `day40` | Autonomous agents: a shared TASK BOARD teammates pull from instead of being hand-fed work. `.mini-agent/tasks/<id>.json` files (on disk = observable), `Task {status, owner, blockedBy}` with `blockedBy` resolved by id OR subject; `create_task`/`list_tasks`/`claim_task`/`complete_task` tools, and `claimTask` is a synchronous read-check-write so two scanners can't both win the same task — no lock, same single-event-loop argument as the mailbox. The teammate IDLE loop now polls inbox-first (shutdown priority) then auto-claims the next ready task (`claimNextAvailable`) and resumes WORK; the lead lays out a dependency graph and teammates respect it (schema → routes → tests), `/tasks` shows the board; `npm test` 574 cases |
 
 ## 📖 配套教程(免费开源)
 
-**[完整教程在 docs/ 目录](docs/README.md)**——主线 + 进阶共十几章,每章对应一个 tag:为什么这么写、Claude Code 原版怎么做、你会踩的坑。每章五段结构:做完长什么样 → 跟着写 → 为什么 → 坑 → 小结。
+**[完整教程在 docs/ 目录](docs/README.md)**——主线 + 进阶共三十多章,每章对应一个 tag:为什么这么写、Claude Code 原版怎么做、你会踩的坑。每章五段结构:做完长什么样 → 跟着写 → 为什么 → 坑 → 小结。
 
-想看更深的「为什么」:公众号的 **Harness Engineering 拆书系列**(30+ 篇)逐章拆了《Harness Engineering:从 Claude Code 看 AI 编码工程》——本仓库每个设计决策背后的原理、数据和事故案例都在那里。公众号 @(待填)。
+想看更深的「为什么」:公众号的 **Harness Engineering 拆书系列**(30+ 篇)逐章拆了《Harness Engineering:从 Claude Code 看 AI 编码工程》——本仓库每个设计决策背后的原理、数据和事故案例都在那里。公众号 @老李是谁。
 
 如果这个项目帮你搞懂了 agent,点个 ⭐ 就是最大的支持。

package/dist/agent.js

@@ -15,15 +15,21 @@ import { connectMcpServers } from "./mcp.js"; // external tool servers (MCP)
 import { Judge } from "./judge.js"; // optional LLM permission classifier
 import { isPlanMode, setPlanMode } from "./permissions.js"; // plan mode: research-only until the user approves a plan
 import { undoLast, clearUndo, sessionChanges } from "./undo.js"; // /undo + /diff: take back, or review, this session's writes
+import { clearTodos, getTodos, renderTodos } from "./todos.js"; // the agent's plan: /todos to view, cleared with the conversation
 import { renderDiff } from "./diff.js"; // show what /undo put back / what /diff changed, reusing the Day 21 diff renderer
 import path from "node:path"; // shorten paths for the /diff summary
 import { expandMentions } from "./mentions.js"; // @file mentions: pull referenced files into context (secret files refused)
 import { banner, framedPrompt, formatModelChoices, statusLine, inputRule } from "./ui.js"; // welcome box, prompt, model labels, status line
-import { gitBranch, toggleLastCollapsed, cleanup as tuiCleanup } from "./tui.js"; // git branch for the status line + collapsible output
+import { gitBranch, toggleLastCollapsed, revealReasoning, clearReasoning, cleanup as tuiCleanup } from "./tui.js"; // git branch for the status line + collapsible output + collapsed reasoning (Ctrl+R)
 import { promptSelect, promptForm } from "./menu.js"; // arrow-key approval menu + multi-question form
+import { editLine } from "./editor.js"; // our own line editor (keeps the status footer pinned even when input wraps)
+import { normalizeDroppedPaths } from "./drop.js"; // drag-and-drop: a dropped file's path → a clean absolute path in the input
 import { rememberTool, readMemory, readMemoryTyped, extractMemories, MEMORY_PATH } from "./memory.js"; // long-term project memory + auto-extract
 import { loadSkills, buildSkillTool, findSkill, skillInstructions } from "./skills.js"; // Markdown-as-plugin skills
 import { initCostMeter, DEFAULT_PRICING } from "./cost.js"; // token & cost accounting for /cost
+import { listBackground, hasRunningBackground, killAllBackground } from "./background.js"; // background tasks: /bg view + kill-on-exit (Day 37)
+import { listTeam, resetTeam } from "./team.js"; // agent teams: /team view + reset on clear/resume (Day 38)
+import { boardSummary, resetBoard } from "./board.js"; // task board: /tasks view + reset on clear/resume (Day 40)
 // package.json sits one level above both src/ (dev) and dist/ (built) — same path either way.
 const pkg = createRequire(import.meta.url)("../package.json");
 // ---- CLI arguments --------------------------------------------------------------
@@ -64,12 +70,20 @@ const SESSION_HELP = `commands:
   /memory    show the durable facts the agent remembers about this project
   /cost      tokens, cache hit rate and estimated spend this session (local)
   /plan      toggle plan mode — research-only; the agent presents a plan you approve before any change
+  /todos     show the agent's current task plan (it maintains one with todo_write on multi-step work)
+  /bg        list background tasks this session (run_bash_background) and their status
+  /team      list the agent team (spawn_teammate): each teammate's role, status, and pending inbox
+  /tasks     show the shared task board (create_task/claim_task): each task's status, owner, and dependencies
   /undo      revert the most recent file write (write_file / edit_file) this session
   /diff      show every file changed this session, as a diff from where it started
   /resume    list recent sessions in this project and continue one of them
   /skills    list the reusable skills available in this project
   /skill <name>  run a skill yourself (works even for user-only skills)
-  exit       leave (Ctrl+C at the prompt does the same)`;
+  exit       leave (Ctrl+C at the prompt does the same)
+
+keys (at the prompt):
+  Ctrl+R     reveal the model's thinking for the last answer (collapsed behind a spinner by default)
+  Tab        expand/collapse the most recent tool output`;
 // Injected when the user turns plan mode on, so the model knows the rules of the
 // mode it now lives in (the permission gate enforces them regardless).
 const PLAN_MODE_NOTICE = `[plan mode ON] Investigate this request using only read-only tools — read_file, search, and safe read-only shell (ls, cat, git status). Do NOT write files, edit, or run mutating commands; the permission gate will block them. When you have a concrete, ordered plan, call the exit_plan_mode tool with that plan. The user reviews and approves it before you make any change.`;
@@ -127,6 +141,7 @@ async function main() {
     // server that fails to start is skipped, never fatal.
     const disconnectMcp = await connectMcpServers();
     process.on("exit", disconnectMcp); // best-effort cleanup of server subprocesses
+    process.on("exit", killAllBackground); // Day 37: SIGKILL any background job (dev server, slow install) so it never outlives the agent as an orphan
     // The optional LLM permission judge, built once if a settings file enabled it.
     const judge = CONFIG.judge.enabled ? new Judge(client, CONFIG.judge.model || CONFIG.model) : undefined;
     if (judge)
@@ -207,6 +222,7 @@ async function main() {
     // ONE readline interface for the whole session — the task prompt and the
     // permission prompts share it. Two interfaces on one stdin fight each other.
     const rl = readline.createInterface({ input: process.stdin, output: process.stdout });
+    rl.setPrompt(""); // a TTY edits via our editor (not rl), so blank readline's default "> " — it would otherwise flash on resume()
     // Input plumbing: a line QUEUE instead of rl.question(). With piped stdin
     // all lines arrive at once — question() would catch the first and lose the
     // rest. The queue catches every line; asking pops one (or waits for one).
@@ -215,7 +231,11 @@ async function main() {
     const pendingLines = []; // lines that arrived before anyone asked
     const lineWaiters = []; // askers waiting for a line
     let stdinDone = false; // has stdin ended?
+    const history = []; // past prompts, for ↑/↓ recall in the editor
+    let editing = false; // true while our line editor owns stdin (so the running-key handler stands down)
     rl.on("line", (line) => {
+        // Only fires for NON-TTY (piped) input — a TTY goes through editLine, which
+        // doesn't emit readline 'line' events. Piped input has no footer to clear.
         const waiter = lineWaiters.shift(); // is somebody waiting?
         if (waiter)
             waiter(line); // hand the line over directly
@@ -228,12 +248,35 @@ async function main() {
             lineWaiters.shift()("exit"); // wake every waiter with a polite "exit"
     });
     // Get the next line of user input, showing a prompt if we have to wait.
-    const readUserLine = (prompt) => {
+    const readUserLine = async (prompt, footer) => {
         if (pendingLines.length)
-            return Promise.resolve(pendingLines.shift()); // typed-ahead (or piped) line — use it
+            return pendingLines.shift(); // typed-ahead (or piped) line — use it
         if (stdinDone)
-            return Promise.resolve("exit"); // stdin is gone — leave politely
-        rl.setPrompt(prompt); // show the prompt the redraw-safe way
+            return "exit"; // stdin is gone — leave politely
+        if (process.stdin.isTTY) {
+            // Our editor owns the whole input block (prompt + wrapped input + footer),
+            // so the status bar stays pinned below however the line wraps.
+            editing = true;
+            try {
+                const res = await editLine(rl, { prompt, footer, history, onTab: () => toggleLastCollapsed(), onReveal: () => revealReasoning(), transformPaste: normalizeDroppedPaths });
+                if (res.type === "line") {
+                    if (res.value.trim())
+                        history.push(res.value); // remember non-empty entries for ↑/↓
+                    return res.value;
+                }
+                if (res.type === "cancel") {
+                    // Ctrl+C at the prompt: exit the session (standard REPL behavior).
+                    console.log();
+                    process.exit(0);
+                }
+                return "exit"; // eof (Ctrl+D)
+            }
+            finally {
+                editing = false;
+            }
+        }
+        // Non-TTY (piped): the readline line queue, no prompt frame, no footer.
+        rl.setPrompt(prompt);
         rl.prompt();
         return new Promise((resolve) => lineWaiters.push(resolve)); // wait for the next line
     };
@@ -396,9 +439,57 @@ async function main() {
                 messages = [{ role: "system", content: systemMessage }]; // drop everything but the constitution
                 forgetFilesExcept([]); // the file read-state belongs to the conversation — clear it too
                 clearUndo(); // a fresh conversation should not undo the previous one's writes
+                clearTodos(); // the plan belonged to the old task — drop it
+                resetTeam(); // the team belonged to the old task — forget it (mailboxes re-wipe on next use)
+                resetBoard(); // and the shared task board (Day 40)
+                clearReasoning(); // drop the collapsed thinking too
                 sessionId = newSessionId(); // a fresh conversation is a fresh session file
                 console.log(chalk.dim("(history cleared)")); // confirm the reset
                 return true;
+            case "/todos": {
+                const todos = getTodos(); // the agent's current checklist for this task
+                console.log(todos.length ? renderTodos(todos) : chalk.dim("(no plan yet — the agent writes one with todo_write on multi-step tasks)"));
+                return true;
+            }
+            case "/bg": {
+                // The background jobs (run_bash_background) started this session — what's
+                // still running, what finished, and with what exit status.
+                const bg = listBackground();
+                if (!bg.length) {
+                    console.log(chalk.dim("(no background tasks — the agent starts them with run_bash_background for slow commands)"));
+                    return true;
+                }
+                const icon = { running: chalk.yellow("●"), completed: chalk.green("✓"), failed: chalk.red("✗"), killed: chalk.dim("∅") };
+                console.log(chalk.dim(`background tasks this session:`));
+                for (const t of bg) {
+                    console.log(chalk.dim(`  ${icon[t.status]} ${t.id} [${t.status}, ${t.elapsed}s] — ${t.command.slice(0, 70)}`));
+                }
+                return true;
+            }
+            case "/team": {
+                // The agent team (spawn_teammate): who's running, who's done, and how
+                // many messages still sit unread in each one's on-disk inbox.
+                const team = listTeam();
+                if (!team.length) {
+                    console.log(chalk.dim("(no team — the agent forms one with spawn_teammate for large parallel tasks)"));
+                    return true;
+                }
+                // Day 39: a running teammate reports as active (working) or idle (waiting).
+                const icon = { active: chalk.yellow("●"), idle: chalk.cyan("◐"), done: chalk.green("✓"), failed: chalk.red("✗") };
+                console.log(chalk.dim(`agent team this session:`));
+                for (const t of team) {
+                    const inbox = t.pending ? chalk.yellow(` · ${t.pending} unread`) : "";
+                    console.log(chalk.dim(`  ${icon[t.status] ?? "?"} ${t.name} [${t.status}, ${t.elapsed}s] — ${t.role.slice(0, 60)}${inbox}`));
+                }
+                return true;
+            }
+            case "/tasks": {
+                // The shared task board (create_task/claim_task/complete_task): what's
+                // pending, claimed, and done, and who owns what.
+                const summary = boardSummary();
+                console.log(summary === "(the task board is empty)" ? chalk.dim("(no tasks — the lead adds them with create_task; teammates claim them autonomously)") : chalk.dim(`task board:\n${summary}`));
+                return true;
+            }
             case "/stats":
                 console.log(chalk.dim(statsReport())); // local counters, busiest first
                 return true;
@@ -462,6 +553,9 @@ async function main() {
                 sessionId = chosen.id; // keep appending to the resumed session's file
                 forgetFilesExcept([]); // a resumed conversation must re-read files before editing them
                 clearUndo(); // the previous session's writes are not ours to undo
+                clearTodos(); // the resumed task starts without the old session's stale plan
+                resetTeam(); // a resumed conversation starts with no live team
+                resetBoard(); // ...and an empty task board (Day 40)
                 console.log(chalk.dim(`(resumed ${chosen.id} — ${chosen.messages.length} messages; files must be re-read before editing)`));
                 return true;
             }
@@ -519,28 +613,47 @@ async function main() {
     };
     // The session loop: ask → run → render → ask again. This is what makes it
     // a conversation instead of a one-shot command.
-    // Listen for Tab key on raw stdin — toggles expand/collapse of hidden tool output.
+    // Raw-stdin key handling that applies while a task is RUNNING. At the prompt
+    // the line editor owns stdin (and handles its own keys, incl. Tab → collapse),
+    // so we stand down whenever `editing` is true to avoid double-processing.
+    //   Esc → interrupt the current operation and return to the prompt (like the
+    //         first Ctrl+C, but without the "press again to force-quit" escalation).
+    // A LONE Esc is one byte (0x1b); arrow keys etc. are multi-byte escape
+    // sequences starting with 0x1b — the length check tells them apart.
     if (process.stdin.isTTY && typeof process.stdin.setRawMode === "function") {
         process.stdin.on("data", (buf) => {
-            if (buf[0] === 0x09) { // Tab
-                if (!running)
-                    toggleLastCollapsed();
+            if (editing)
+                return; // the editor is in charge of these keys
+            if (buf.length === 1 && buf[0] === 0x1b) {
+                // Esc: stop the current operation, keep the session alive and typeable.
+                if (running && !interrupted) {
+                    interrupted = true; // the loop polls this between steps
+                    controller.abort(); // cancel the in-flight API request
+                    console.log(chalk.yellow("\n(esc — interrupted; back at the prompt)"));
+                }
             }
         });
     }
     while (true) {
-        // Before each prompt, a status line (model · 📁 dir · 🌿 branch · ctx% · $ · time)
-        // then a separator rule, then the "❯". Only in a TTY; piped input is clean.
+        // The input area, Claude-Code style: a top rule, the "❯" you type on, then a
+        // bottom rule + status line (model · 📁 dir · 🌿 branch · ctx% · $ · time)
+        // pinned BELOW the input. readUserLine draws the footer under the prompt and
+        // keeps the cursor on the ❯ line, so the status reads like a real bottom bar.
+        // Only in a TTY; piped input stays clean.
+        let footer;
         if (process.stdin.isTTY) {
             const ctxPct = Math.min(100, Math.round((estimateHistoryTokens(messages) / CONFIG.contextWindow) * 100));
-            console.log("\n" + statusLine(CONFIG.model, path.basename(process.cwd()), gitBranch(), ctxPct, costMeter.cost(), Date.now() - sessionStartedAt));
-            console.log(inputRule());
+            console.log("\n" + inputRule()); // top rule above the prompt
+            footer = inputRule() + "\n" + statusLine(CONFIG.model, path.basename(process.cwd()), gitBranch(), ctxPct, costMeter.cost(), Date.now() - sessionStartedAt); // bottom rule + status, pinned below
         }
-        const line = (await readUserLine(framedPrompt(isPlanMode()))).trim(); // next input, or "exit" on EOF
+        const line = (await readUserLine(framedPrompt(isPlanMode()), footer)).trim(); // next input, or "exit" on EOF
         if (!line)
             continue; // empty line — just re-prompt
-        if (line === "exit" || line === "quit")
+        if (line === "exit" || line === "quit") {
+            if (hasRunningBackground())
+                console.log(chalk.yellow("(stopping background tasks still running — see /bg)")); // Day 37: they're about to be SIGKILLed on exit
             break; // explicit goodbye
+        }
         if (await handleCommand(line))
             continue; // slash commands never reach the model
         // UserPromptSubmit hook: a chance to validate/inject before the model sees
@@ -563,6 +676,7 @@ async function main() {
         if (refused.length)
             console.log(chalk.yellow(`(refused secret file${refused.length === 1 ? "" : "s"}: ${refused.join(", ")})`));
         messages.push({ role: "user", content: augmented + injected }); // the new turn (with any attached files + hook context) joins the shared history
+        clearReasoning(); // Ctrl+R should reveal THIS turn's thinking, not the previous answer's
         running = true; // Ctrl+C now means "interrupt the task"
         interrupted = false; // fresh interrupt state for this task
         controller = new AbortController(); // fresh abort signal for this task

package/dist/background.js

@@ -0,0 +1,160 @@
+// ============ Background tasks (Day 37) ============
+// The problem: run_bash blocks. A command that takes longer than its 30s
+// watchdog (npm install, a full test run, a build) is SIGKILLed before it
+// finishes, and even a 25s one freezes the agent — we sit idle, paying for
+// tokens we aren't using, while a child process churns. The model cannot think
+// and wait at the same time.
+//
+// The fix, mirroring Claude Code: a command can be sent to the BACKGROUND. We
+// spawn it, hand the model a task id immediately, and let the loop keep going.
+// When the command finishes, the loop injects a <task_notification> as the next
+// user turn — so the result rejoins the conversation without ever having blocked
+// it. Long-running servers (npm run dev) that never "finish" are supported too:
+// they stay `running` forever and the model polls them with bash_output.
+//
+// Why no extra thread / lock (unlike the Python reference): Node is single-
+// threaded with an event loop. `spawn` is already async — stdout/stderr arrive
+// as events while the main loop runs. There is no shared-memory race to guard,
+// so the whole "daemon thread + mutex" dance collapses into plain callbacks.
+import { spawn } from "node:child_process";
+const tasks = new Map(); // id → task, for the whole session
+let seq = 0; // increments per task — seeds the id
+const OUTPUT_CAP = 80_000; // chars of output retained per task (keep the tail)
+const SUMMARY_TAIL = 1_500; // chars of the tail shown inline in a notification
+// Append output to a task, keeping only the most recent OUTPUT_CAP chars. We
+// keep the TAIL, not the head: an exit error or a "server ready on :3000" line
+// lands at the end, and that is what the model needs to react to. `produced`
+// counts everything ever written so the poll cursor stays correct across trims.
+function append(task, chunk) {
+    task.produced += chunk.length;
+    task.output += chunk;
+    if (task.output.length > OUTPUT_CAP)
+        task.output = task.output.slice(-OUTPUT_CAP);
+}
+// Start a command in the background. Returns its task id immediately — the
+// command keeps running after this returns. Note the deliberate omissions vs
+// run_bash: NO 30s timeout (the whole point is to outlive it) and NO turn
+// AbortSignal (Ctrl+C interrupts the current turn, but a background job is meant
+// to survive into later turns; only killAll, on session exit, stops them).
+export function startBackground(command) {
+    const id = `bg_${++seq}`;
+    const child = spawn(command, { shell: true, stdio: ["ignore", "pipe", "pipe"] }); // shell:true → bash semantics, same as run_bash
+    const task = {
+        id,
+        command,
+        child,
+        status: "running",
+        exitCode: null,
+        output: "",
+        produced: 0,
+        readAt: 0,
+        notified: false,
+        startedAt: Date.now(),
+        endedAt: null,
+    };
+    tasks.set(id, task);
+    child.stdout?.on("data", (d) => append(task, d.toString())); // stream output into the buffer as it arrives
+    child.stderr?.on("data", (d) => append(task, d.toString()));
+    child.on("error", (err) => {
+        // The process could not even start (e.g. command not found). Record it as a
+        // failure with the reason, so the notification explains what happened.
+        append(task, `\n[could not start: ${err.message}]`);
+        if (task.status === "running") {
+            task.status = "failed";
+            task.exitCode = null;
+            task.endedAt = Date.now();
+        }
+    });
+    child.on("close", (code, signal) => {
+        if (task.status !== "running")
+            return; // already terminal (killAll set `killed`, or an error fired) — don't overwrite
+        task.exitCode = code;
+        task.status = code === 0 ? "completed" : "failed";
+        if (signal)
+            task.status = "killed"; // killed by a signal, not its own exit
+        task.endedAt = Date.now();
+    });
+    return id;
+}
+// Collect notifications for every task that has FINISHED since the last call.
+// Each task is reported exactly once (the `notified` flag), so calling this
+// every round is safe: a finished task surfaces once, then never again. Returns
+// "" when there is nothing to report. The <task_notification> shape is taken
+// straight from the tutorial reference so the format is consistent across the
+// course.
+export function pendingNotifications() {
+    const done = [];
+    for (const task of tasks.values()) {
+        if (task.status === "running" || task.notified)
+            continue; // still going, or already reported
+        task.notified = true; // report it once
+        const secs = task.endedAt ? Math.round((task.endedAt - task.startedAt) / 1000) : 0;
+        const tail = task.output.slice(-SUMMARY_TAIL).trim() || "(no output)";
+        const more = task.output.length > SUMMARY_TAIL ? " (truncated — use bash_output for the full log)" : "";
+        done.push(`<task_notification>\n` +
+            `  <task_id>${task.id}</task_id>\n` +
+            `  <status>${task.status}</status>\n` +
+            `  <command>${task.command}</command>\n` +
+            `  <exit_code>${task.exitCode ?? "n/a"}</exit_code>\n` +
+            `  <elapsed>${secs}s</elapsed>\n` +
+            `  <summary>${tail}${more}</summary>\n` +
+            `</task_notification>`);
+    }
+    if (!done.length)
+        return "";
+    // The framing line matters: without it the model can mistake the XML for
+    // something it must echo. Tell it plainly that a job it started has finished.
+    return `A background task you started has finished. Read the notification, then continue.\n\n${done.join("\n\n")}`;
+}
+// Read a background task's current state for the bash_output tool. Returns only
+// the output PRODUCED SINCE THE LAST POLL (like Claude Code's BashOutput), so
+// the model can watch a long job without re-reading the whole log each time.
+// `full: true` returns the entire retained buffer instead.
+export function readBackground(id, full = false) {
+    const task = tasks.get(id);
+    if (!task) {
+        const known = [...tasks.keys()];
+        return `[error] No background task "${id}". ${known.length ? `Known tasks: ${known.join(", ")}.` : "No background tasks have been started this session."}`;
+    }
+    let chunk;
+    if (full) {
+        chunk = task.output; // the entire retained buffer
+    }
+    else {
+        const fresh = task.produced - task.readAt; // chars produced since last poll
+        // Trimming may have dropped older chars; never read past what we still hold.
+        chunk = task.output.slice(Math.max(0, task.output.length - fresh));
+    }
+    task.readAt = task.produced; // either way, the model has now seen up to here — next poll shows only newer output
+    const head = `[${task.id}] ${task.status}${task.exitCode !== null ? ` (exit ${task.exitCode})` : ""} · ${task.command}`;
+    // Be precise about an empty body: nothing ever vs. nothing NEW since the last
+    // poll (the model already read it) vs. nothing new yet on a still-running job.
+    const body = chunk.trim() || (task.produced === 0 ? "(no output)" : task.status === "running" ? "(no new output yet)" : "(no new output)");
+    return `${head}\n${body}`;
+}
+// A one-line-per-task snapshot for the /bg command. Newest first.
+export function listBackground() {
+    return [...tasks.values()]
+        .sort((a, b) => b.startedAt - a.startedAt)
+        .map((t) => ({
+        id: t.id,
+        status: t.status,
+        command: t.command,
+        elapsed: Math.round(((t.endedAt ?? Date.now()) - t.startedAt) / 1000),
+    }));
+}
+// True if any task is still running — the REPL uses this to warn before exit.
+export function hasRunningBackground() {
+    return [...tasks.values()].some((t) => t.status === "running");
+}
+// Kill every still-running task. Called on session exit so a forgotten dev
+// server or a slow install does not outlive the agent as an orphan.
+export function killAllBackground() {
+    for (const task of tasks.values()) {
+        if (task.status !== "running")
+            continue;
+        task.child.kill("SIGKILL");
+        task.status = "killed";
+        task.endedAt = Date.now();
+    }
+}