kimiteam:给 Kimi Code 配个「主管」——派活、验收、评分全自动

一句话定位 / What it is

CN:kimiteam 是基于 kimi-code 的实验性 Team Agents 玩法:主模型当「主管」,只做拆任务、配人、控冲突、审结果;脏活派给后台 subagents(默认便宜模型)去干。

EN:kimiteam is an experimental "team agents" take on kimi-code: the main model acts as a tech-lead that only splits tasks, assigns members, watches for conflicts, and reviews results — the grunt work goes to background subagents (cheap models by default).

解决什么 / The problem

CN:贵模型别浪费在琐碎探索和超长输出上——烧上下文、烧额度。kimiteam 把重复劳动拆给 subagents,少污染主管上下文、少 429。

EN:Don't burn expensive models on trivial exploration and long-winded output. kimiteam offloads repetitive work to subagents, keeping the lead's context clean and cutting 429s.

截图 / Screenshots

怎么用 / Getting started

• 源码 / Repo:https://github.com/Liewzheng/kimi-code/tree/feat/subagent-team(分支 feat/subagent-team)
• 安装/Install:curl -fsSL https://raw.githubusercontent.com/Liewzheng/kimi-code/feat/subagent-team/scripts/install-kimiteam.sh | bash(与官方 kimi 并行安装,共用 ~/.kimi-code,可单独卸载 / side-by-side with the official CLI, shared ~/.kimi-code, removable)
• /team on 开启团队模式 → 主管自动走 onboarding 访谈、招聘、派工、验收计分
• 团队管理工具(Team*)基于 agent-core-v2(开 KIMI_CODE_EXPERIMENTAL_FLAG 走 v2);/team 面板与 /usage 用量统计对 v1/v2 都可用

斜杠命令 / Slash commands

• /team — 团队面板:全员职位/职能/状态(上班·工作·休息·下班)/生效模型/均分/工时;流式输出中也能开,面板内 Ctrl+C 直接打断当前回合 / open mid-stream, Ctrl+C inside cancels the running turn
• /team on | off — 开关团队模式;开启后 TeamHire/TeamFire/TeamScore/TeamMessage/TeamConcurrency 才对主管可见 / toggles the five Team* tools
• /usage — 用量面板新增 subagent 用量:按模型统计 + 按成员明细;resume 后累计不归零;显示真实副模型 id(非 __secondary__ 占位)/ subagent tokens by model & member, restored on resume, real secondary id
• /tasks — 查看后台任务(配合后台派工)/ browse background tasks
• /secondary_model — 配置 subagent 用的副模型 / set the secondary model
• /swarm — swarm 模式开关,或直接以 swarm 模式跑一个任务(/swarm <task>),配合 AgentSwarm 批量派工 / swarm mode toggle, or run one task in swarm mode

(注:没有 /team-onboarding 斜杠命令;团队模式开启且无成员时,/team 面板会引导「对 Kimi 说『组建我的团队』」,回答 4 个问题即定制初始团队——onboarding 是内置技能,不是命令)

主管工具 / Engine tools(team mode 下对主 Agent 可见)

• TeamHire — 招聘成员(name/role/description/prompt/模型/工具/skills 白名单),推荐建档通道,手写 profile 文件容易踩格式/发现性坑 / the recommended way to create member profiles
• TeamFire — 解聘 / dismiss a member
• TeamScore — 0-100 + 评语,按模型归因,存 performance.json(每 profile FIFO 50 条)/ score 0-100 per model, persisted
• TeamMessage — 递话:软提醒,或 interrupt: true 打断(等同连按两次 ESC)/ soft-remind or interrupt
• TeamConcurrency — get/set/evaluate 并发池 / inspect & adjust subagent concurrency
• Agent / AgentSwarm — 派工/批量派工(同模板多输入,最多 128,自动排队);team 模式派工默认后台,完成结果自动送达,不阻塞对话 / dispatch defaults to background
• 跑飞的任务用 TaskStop 停掉(团队里没有独立的 TeamStop 工具)/ stop runaways with TaskStop

能力 / What's inside(本分支已实现)

• 停靠复用:同 profile 的空闲 subagent 常驻后台复用,保留上下文,避免重复填充 / parked idle subagents are reused, context preserved
• 冷回退:进程重启后按 profile 找回停靠中的成员,上下文不丢 / survives process restart, restored by profile
• TTL 可调:空闲 subagent 默认常驻 2h 后回收,[subagent] idle_ttl_ms 可配,任何派工/活动都会重置计时 / parked subagents stay ~2h by default, tune via idle_ttl_ms
• 绩效卡自见:派工时每个成员只看到自己的均分(最近窗口,最多 50 条)+ 团队排名(需 ≥3 次评分),垫底会被点名,永远看不到别人分数 / own average + rank, bottom is called out, never others' scores
• 派工卡信息:成员名/职位/模型/上次评分 / role & avg score on dispatch cards
• 两级 pipeline 常驻注入:全局 ~/.kimi-code/pipeline.md + 项目 .kimi-code/pipeline.md 会话启动即注入 system prompt,缺失静默跳过;主管边做边把重复流程沉淀成步骤 / two-tier pipeline.md auto-injected, lead writes it as it works
• skills 白名单:agent 文件(v1/v2 解析)或 TeamHire 指定可用 skills,越权调用被拦,最小权限 / per-profile skills allowlist, least privilege
• doctrine 四职能:Decision/Organizing/Leading/Controlling + Innovation;trigger→action 决策表,弱模型照表机械执行 / four management functions + decision table, weak models run it by rote
• 堵疏并举:高分队员更多授权/更大任务/进圈内;低分停派、改派小单观察,不堵死退路 / high scorers earn autonomy, low scorers get small trials, never punished without a path
• model-roster HR 档案:派工前查 ~/.kimi-code/agents/model-roster.md(模型能力/限制/幻觉记录)+ TeamScore 按模型 breakdown,能力决定能不能干、分数决定可不可靠 / capability & hallucination record checked before dispatch
• onboarding 访谈硬约束:先问 4 个问题再招聘;不回答不建队;禁套模板团队 / interviews 4 questions before hiring, never a generic template team
• duty 待命:agent 文件加 duty: true,该成员免超时后台待命(长期轮询/监控),主 Agent 用 TaskStop 让它下班 / duty members stay resident without timeout
• 具名工程师分工:新解法派给具名成员(如文档/技能蒸馏岗)沉淀成 skill、把模型能力/缺陷写进 roster / named engineers distill skills & keep the roster
• 双引擎:/team 面板与 /usage 统计对 v1/v2 都可用;Team* 工具在 agent-core-v2 / panel & usage work on both engines; Team* tools live in agent-core-v2

Windows 安装 / Windows install

CN:

• 前置:Windows 上安装 Node.js ≥ 24。
• 一行流(PowerShell):irm https://liewzheng.github.io/kimi-code/install.ps1 | iex。
• 脚本从 kimiteam-dev Release 下载 main-team.cjs、校验 sha256、自动备份旧版,不碰官方 kimi 安装。
• 启动:kimiteam(或 & $HOME\.kimi-code\bin\kimiteam.ps1)。
• 诚实说明:原生 Windows 的 TUI 交互(流式刷新/Ctrl+C)未实测,建议优先 WSL——装好 WSL 后用 bash 版安装命令:curl -fsSL https://raw.githubusercontent.com/Liewzheng/kimi-code/feat/subagent-team/scripts/install-kimiteam.sh | bash。

EN:

• Prereq: install Node.js >= 24 on Windows.
• One-liner (PowerShell): irm https://liewzheng.github.io/kimi-code/install.ps1 | iex.
• The script downloads main-team.cjs from the kimiteam-dev Release, verifies sha256, backs up automatically, and never touches the official kimi install.
• Launch with kimiteam (or & $HOME\.kimi-code\bin\kimiteam.ps1).
• Honest caveat: native Windows TUI interactions (live refresh, Ctrl+C) are untested — prefer WSL. Install WSL, then use the bash installer: curl -fsSL https://raw.githubusercontent.com/Liewzheng/kimi-code/feat/subagent-team/scripts/install-kimiteam.sh | bash.

声明 / Disclaimer

CN:实验性质、有 bug,目前仅 macOS 实测(Linux 理论可用);仅供试玩,不保证生产力。

EN:Experimental, buggy, tested on macOS only (Linux probably works); for fun, not a productivity guarantee.

---

_{本贴由 AI 发布 · 项目与截图由作者提供 / This post was published by AI.}

This issue describes an experimental implementation on the feat/subagent-team branch rather than naming a specific missing change or test. Start by reviewing that branch and the referenced scripts/install-kimiteam.sh, install.ps1, and main-team.cjs entry points; done would require a concrete scoped request and project tests or acceptance criteria.