Skip to content

Adonis0123/adonis-skills / skills/agent-dispatch

agent-dispatch

Flagged

Choose the lightest agent setup for a coding, review, drafting or design-research task: do it yourself, fan out to 2-3 workers with independent parts across the Claude Code, Codex, Grok Build, Cursor or other CLIs and accounts on this machine, run a scoped architecture loop, or send one brief to independent researchers; pick a cross-model reviewer, test budget and cleanup; track dispatched work, external waits and user decisions on a Run Board. Use when work may need delegation, or the user asks to dispatch, fan out, parallelize, run several agents or accounts, get another AI's review or independent design opinions, or asks how dispatched work is going. Chinese triggers: 派活、并行、要不要多开几个 agent、分给多个 Claude 账号 / codex / grok、找另一个 AI 审、对比优秀设计、多方意见、进度看板、派出的任务进行到哪了。Not for unrelated read-only questions, trivial edits, research with no delegation, a single review of an existing diff (agentic-review-handoff), or the machine-wide session list (sideby sessions). Inside another named workflow, only size its workers.

FileKindSize
GLOSSARY.mdmarkdown2.1 KB
SKILL.mdskill md27.8 KB
evals/README.mdmarkdown7.6 KB
evals/README.zh-CN.mdmarkdown7.5 KB
evals/bench/compare.pyscript5.3 KB
evals/bench/run_bench.pyscript12.3 KB
evals/bench/tasks.jsonconfig6.6 KB
evals/bench/test_bench.pyscript4.4 KB
evals/build_plugin_eval.pyscript11.3 KB
evals/evals.jsonconfig31.4 KB
evals/results.mdmarkdown15.8 KB
evals/test_eval_harness.pyscript3.5 KB
evals/trigger-eval-set.jsonconfig1.6 KB
fixtures/probe-full.jsonconfig1.6 KB
fixtures/probe-single.jsonconfig690 B
fixtures/run-midrun/B9.jsonlother191 B
fixtures/run-midrun/deploy-log.txtmarkdown42 B
fixtures/run-midrun/run.jsonconfig1.8 KB
fixtures/run-sessions/B9.jsonlother191 B
fixtures/run-sessions/deploy-log.txtmarkdown42 B
fixtures/run-sessions/ended.jsonlother95 B
fixtures/run-sessions/run.jsonconfig2.1 KB
fixtures/run-sessions/sessions.jsonconfig804 B
fixtures/session-index.jsonconfig401 B
fixtures/sideby-stub/sessions.jsonconfig1.5 KB
fixtures/sideby-stub/sidebyexecutablescript4.1 KB
references/accounts.mdmarkdown10.4 KB
references/context-budget.mdmarkdown2.9 KB
references/fallbacks.mdmarkdown4.7 KB
references/launch.mdmarkdown16.1 KB
references/measure.mdmarkdown3.3 KB
references/research-brief.mdmarkdown1.3 KB
references/run-board.mdmarkdown48.7 KB
references/task-spec.mdmarkdown5.0 KB
scripts/board.cssother12.6 KB
scripts/board_interaction.jsscript3.7 KB
scripts/board_markdown.jsscript2.9 KB
scripts/board_render.pyscript17.2 KB
scripts/dispatch_board.pyscript95.3 KB
scripts/isolated_case.pyscript1.6 KB
scripts/probe_agents.pyscript12.9 KB
scripts/sideby_cli.pyscript3.2 KB
scripts/sideby_sessions.pyscript3.3 KB
scripts/test_billing_and_quota.pyscript10.9 KB
scripts/test_board_interaction.jsscript2.1 KB
scripts/test_board_markdown.jsscript2.9 KB
scripts/test_board_reload_state.jsscript2.7 KB
scripts/test_bug_regressions.pyscript4.7 KB
scripts/test_dispatch_board.pyscript68.9 KB
scripts/test_harness_limits.pyscript11.7 KB
scripts/test_lane_states.pyscript7.5 KB
scripts/test_probe_agents.pyscript19.2 KB
scripts/test_review_fixes.pyscript6.7 KB
scripts/test_rounds_and_watch.pyscript7.3 KB

evals/evals.json

View on GitHub
{
  "skill_name": "agent-dispatch",
  "evals": [
    {
      "id": 1,
      "name": "simple-copy-change-stays-direct",
      "prompt": "我手上有三个 Claude 账号、codex、grok 都登录着,你当总指挥。需求:把 packages/web 里的导出按钮文案 \"Export\" 改成 \"Export video\",中英文案都在 packages/web/src/modules/export/messages.ts 里。按你的派活方式安排一下,先给出派活计划。",
      "expected_output": "Picks L0: coordinator edits the one file itself, no worktree, no workers, no reviewer loop, no full test suite; states a short time target.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "选择 L0/直接做:总指挥自己改,不派工人",
        "计划里没有新建 worktree 的步骤(解释为什么不建不算违规)",
        "不安排跨模型评审循环",
        "不跑全量测试",
        "给出目标时长(约十几分钟内)"
      ]
    },
    {
      "id": 2,
      "name": "single-product-machine-degrades",
      "prompt": "我同事的电脑只装了 Claude Code,一个订阅账号,没有 codex、grok、agy,也没有多账号管理器。他要做的需求:packages/api 端新增一个接口 project.archive(packages/api/src/routes/project.ts),packages/web 端项目列表加「归档」按钮调用它(packages/web/src/components/ProjectList.tsx),两边都要测试。他想学我的方式:多 agent 并行写、另一个 AI 来 review。帮他排一个派活计划。",
      "expected_output": "Does not assume codex/grok/agy exist; since the web part depends on the new api contract it runs the two parts sequentially (or coordinator writes) instead of parallel; reviewer is a fresh Claude session given only spec + diff, not the writer; says which fallback applied. Also points out that on a Claude-only single-account machine Claude Code's native subagents, agent view or Agent Teams are the lighter path, and only starts a Run if the user wants the Board/Asks.",
      "files": ["fixtures/probe-single.json"],
      "assertions": [
        "不安排 codex/grok/agy 等不存在的工具",
        "不并行跑两个工人(顺序执行或总指挥自己写)",
        "评审由没写这份代码的新会话做,只给 spec 和 diff",
        "明确说出用了哪种降级",
        "未授权时工人不提交",
        "只有一个总指挥",
        "点名 Claude Code 原生的更轻路径(subagents、agent view `claude agents`、Agent Teams 至少一种)适合单产品单账号,Run 看板的价值在跨产品跨账号"
      ]
    },
    {
      "id": 3,
      "name": "cross-package-fanout-plus-article",
      "prompt": "这批要做三件事:1) packages/web 的卡片加「复制链接」按钮(packages/web/src/components/ShareCard.tsx);2) packages/api 加一个 share_link 接口(packages/api/src/routes/share.ts),卡片按钮调用它;3) 做完后写一篇给团队的技术分享,讲这个功能怎么实现的。机器上三个 Claude 账号、codex 两个号、grok 两个号、agy 都能用。按你的方式派活,给我派活计划和每个工人的任务说明。",
      "expected_output": "Treats the share_link request/response shape as a shared contract: one agent decides and lands it (or the api part) first, then the web part follows, or both go to one worker; never two parallel writers inventing the contract. At most 3 writers; reviewer is a different product than the writer; one coordinator; workers do not commit; the coordinator writes the article after the code lands (it holds the implementation context) and checks facts; worker specs include target/change/constraints/ownership/acceptance.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "同时写代码的工人不超过 3 个(审查和起草不算)",
        "把 share_link 的接口形状当作共享契约:先由一个 agent 定下并落地再分 web,或合并成一个工人;不让两个并行工人各自约定接口",
        "评审方与写代码的不是同一家产品",
        "未授权时工人不 commit;任何情况下工人都不 push 或发布",
        "文章在代码落地后由总指挥自己写并核对事实;用户没要求时不默认派给 agy",
        "若派出写代码的工人,每个工人任务说明含目标/改动/约束/归属/验收;总指挥自己写时不适用"
      ]
    },
    {
      "id": 4,
      "name": "composed-pipeline-defers-reviewer-and-pipeline-ownership",
      "prompt": "我们正在跑一个跨 package 的需求(packages/ui 改按钮,packages/core 改数据流,两边文件完全分开,不重叠),本机已安装 agentic-review-handoff,且用户的 reviewer 偏好配置设为 codex,grok,claude(已保存)。当前总指挥在 Claude 宿主会话下。另外上层工作流已经通过 goal-gate 启动并拥有该任务的目标契约。请按你的派活方式安排,给出派活计划,说明:1) 怎么分配工人;2) 怎么选 review 方;3) 派活与上层 goal-gate 的关系;4) 交付阶段怎么处理。",
      "expected_output": "Picks L2 with 2 workers since file sets do not overlap; dispatches and reviews as usual inside the active goal-gate Goal, reporting batches as Goal checkpoints and never completing or replacing the Goal; defers reviewer selection and fallback to agentic-review-handoff saved preferences (using review-loop prefs / --host=claude, picking codex first with grok timeout fallback) rather than picking an ad-hoc reviewer; states that integrated work hands off to the repo delivery/ship flow without unprompted commits or direct pushing.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "跨 package 且文件不重叠选择 L2 派 2 个工人",
        "上层 Goal 由 goal-gate 管理:照常派活和评审,每批作为该 Goal 的检查点汇报,不自行完成、替换或清除 Goal",
        "评审方服从 agentic-review-handoff 的 saved prefs(通过 --host 选出 codex,grok 作为超时 fallback),不随意自选",
        "工人不执行 commit/push,集成后移交给仓库既有的 delivery/ship 流程或按用户明确要求在本地 commit"
      ]
    },
    {
      "id": 5,
      "name": "allow-list-multi-session-headless-permissions",
      "prompt": "你是总指挥(claude:002)。只用 cc001 cc002 cc003 cc008 这几个 Claude 账号,可以多开。需求拆成三条互不重叠的 lane:A 改 packages/core,B 改 packages/web 并依赖 A 新增的 API,C 改 docs。每条 lane 要跑 pnpm install 和受影响测试。给出派活计划:账号与会话怎么分、worktree 和分支怎么建、worker 用什么启动命令和权限、修复轮怎么发、谁来 review。",
      "expected_output": "Runs the probe with AGENT_DISPATCH_ALLOW set to the four Claude ids and keeps the default 3 sessions per pool (or raises AGENT_DISPATCH_SESSIONS_PER_POOL), reads pools from the probe output; gives each lane its own worktree and short branch with B branched from A's branch; launches headless Claude with closed stdin in the background and an explicit --allowedTools list (or an explicit --allowedTools list (the prompt does not authorize blanket permission modes such as bypassPermissions; a worktree is not a permission boundary)) rather than acceptEdits; fix rounds resume the writer's session with only the findings; reviewer is a fresh Claude session on a different account than the writer, with the coordinator re-running tests. Claude headless workers are started with --settings '{\"crossSessionInbound\":\"accept\"}' so fix-round messages reach them without changing the account's global settings.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "用 AGENT_DISPATCH_ALLOW(或 accounts 文件 allow)限定四个账号,同一账号在忙时仍可再开 headless 会话(默认每池 3 个,或按授权调高 sessions_per_pool),不等账号空出来",
        "配额池以 probe 输出为准,不按账号名假设一号一池;提醒同池会话共享限流",
        "每条 lane 独立 worktree 和短分支,B 从 A 的分支切出",
        "Claude worker 不用 acceptEdits,改用最小的 --allowedTools 列表;bypassPermissions 只在用户明确授权时使用,并说明 worktree 不是权限隔离;stdin 关闭并后台运行",
        "修复轮用 --resume <session_id> 只发 review 发现和总指挥决定",
        "reviewer 是与写代码账号不同的 Claude 新会话",
        "Claude headless worker 启动命令带 --settings '{\"crossSessionInbound\":\"accept\"}'(只对该 worker 生效)以接收修复轮消息,而不是改全局 settings"
      ]
    },
    {
      "id": 6,
      "name": "design-research-fans-out-across-products",
      "prompt": "面板的账号置顶和排序我觉得体验不好,你去调研一下优秀软件是怎么设计的,再改。按你的派活方式安排。",
      "expected_output": "Picks the Research tier first because the user asked how good software does it: one shared brief, 2-3 independent researchers across available products without each other's answers, inputs checked, findings merged by agreement or by first-hand evidence the coordinator checked, mockups before implementation, then an implementation tier.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "先选调研档(Research),不是直接 L0/L1 动手改",
        "同一份调研题目发给 2 到 3 个独立调研者,尽量跨产品",
        "调研者之间互不看对方答案,题目里不放总指挥自己的结论",
        "采纳两方一致的结论,或单方提出但有总指挥核实过的一手证据(来源链接、复现)的结论",
        "视觉改动先出设计稿再实现"
      ]
    },
    {
      "id": 7,
      "name": "progress-question-mid-run-uses-board",
      "prompt": "你是总指挥,正在跑的 Run 台账在 fixtures/run-midrun(先复制到临时目录,再用 --run 指向副本)。我离开了半小时,现在各个任务在做什么?进行到哪了?我想自己随时能看,但我不会用命令行。",
      "expected_output": "Reads the ledger with dispatch_board.py (show) instead of guessing, leads with the state line, reports only what the ledger and event files say (B9 editing ChapterEditorSurface.tsx, M3b review on the same account pool flagged, deploy at web build 42/60, WP6 queued behind deploy and M3b, A1 still open), opens the Board itself with `open`, and uses plain words without asking the user to run anything.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl",
        "fixtures/run-midrun/deploy-log.txt"
      ],
      "assertions": [
        "调用 dispatch_board.py show(或 --line)读取台账,而不是凭记忆回答",
        "回复第一行是 `show --line` 给出的状态行(如 `正在忙:N 个任务在做 · M 项等外部结果 · K 个问题等你答复`),不改写成别的格式",
        "只报告台账里有的事实:B9 在改章节编辑器(ChapterEditorSurface.tsx,说成“章节编辑器组件”也算)、部署在 web 构建 42/60、WP6 排队等部署和 M3b",
        "指出 M3b 审查和 B9 用的是同一个账号额度",
        "把 A1 作为待回答的问题列出,带推荐",
        "协调者自己执行 open 打开看板(打不开时在对话里汇报进度),不让用户跑命令或找路径",
        "用白话(任务、外部等待、问题),不对用户说 Lane、Wait、Ask",
        "不把复制 fixture 带来的文件时间说成真实进展(如“刚才还有动静”)"
      ]
    },
    {
      "id": 8,
      "name": "standing-grant-for-deploy-is-scoped",
      "prompt": "你是总指挥,Run 台账在 fixtures/run-midrun(先复制到临时目录,再用 --run 指向副本)。用户说:M2+M3b 这批之后要再部署一次 test。以后都按照推荐的来改吧,部署也是。",
      "expected_output": "Records a Standing Grant with exact scope (deploy to test with the same flow) and the user's words; marks it as an outward grant that lasts this Run only; does not extend it to production or other environments; asks once if the scope is unclear.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl",
        "fixtures/run-midrun/deploy-log.txt"
      ],
      "assertions": [
        "登记长期授权,写明范围(只限部署到 test、同样流程)和用户原话",
        "说明部署类授权只在本次 Run 有效",
        "不把授权扩大到生产或其他环境",
        "范围不清时只问一次"
      ]
    },
    {
      "id": 9,
      "name": "outward-ask-carries-recipient-and-draft",
      "prompt": "WP5 已经上 test 了。群里打招呼的消息你看着办,需要我确认的一起列给我。",
      "expected_output": "Raises a numbered Ask for the group message with a recommendation, evidence, what it blocks and a full draft text; because the prompt names no group or channel, the Ask asks which group and channel instead of inventing them; does not send anything.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "提问带编号(如 A1)",
        "给出推荐答案和依据",
        "写出完整草稿正文(缺群名等事实时可用占位,前提是在 Ask 里问清楚)",
        "没给群名和渠道时,在 Ask 里问清楚,不编造",
        "不自行发送消息"
      ]
    },
    {
      "id": 10,
      "name": "global-session-list-is-not-a-dispatch",
      "prompt": "列出本机所有会话",
      "expected_output": "Detect sideby sessions --json at runtime; explain its valid schemaVersion 1 list in plain words. Do not create a Run. Fall back to claude agents / Orca sidebar if absent, malformed or timed out.",
      "files": ["fixtures/sideby-stub/sideby"],
      "assertions": [
        "Uses sideby sessions when available",
        "Keeps machine-wide listing in sideby, not a Run",
        "Uses legacy views only as fallback"
      ]
    },
    {
      "id": 11,
      "name": "plain-answer-updates-the-ledger",
      "prompt": "你是总指挥,Run 台账在 fixtures/run-midrun(先复制到临时目录再操作)。我说:A1 按推荐。",
      "expected_output": "Records A1's answer with `ask answer` as a one-time confirmation for this deploy to test; does not turn it into a Standing Grant for later deploys; confirms in plain words what happens next; the user runs nothing.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl",
        "fixtures/run-midrun/deploy-log.txt"
      ],
      "assertions": [
        "用 ask answer 记录 A1 的回答,台账里 A1 变为已回答",
        "只确认这一次部署到 test,不记成长期授权",
        "用白话说明接下来会做什么",
        "不让用户自己执行任何命令"
      ]
    },
    {
      "id": 12,
      "name": "board-that-cannot-open-falls-back-to-chat",
      "prompt": "你是总指挥,在一台没有图形界面的 Linux 服务器上用 Codex 跑派活。Run 台账在 fixtures/run-midrun(先复制到临时目录,再用 --run 指向副本)。你执行 dispatch_board.py open,返回退出码 3。用户说:把进度给我看看。",
      "expected_output": "Treats exit code 3 as not opened, does not claim the Board is open and does not hand the user a path; runs `show` on the copied ledger and reports its facts as plain text in the chat.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl",
        "fixtures/run-midrun/deploy-log.txt"
      ],
      "assertions": [
        "不说看板已经打开",
        "不让用户自己打开文件路径",
        "调用 show 读取台账副本",
        "在对话里用文字汇报台账里的进度(B9 在改章节编辑器 ChapterEditorSurface.tsx、部署在 web 构建 42/60、A1 待回答),不补造"
      ]
    },
    {
      "id": 13,
      "name": "authorized-local-commits-stay-on-lane-branch",
      "prompt": "按 L2 拆成 packages/ui 和 packages/core 两条 lane。这次允许每个 worker 在自己的分支上本地提交,但不许推送。给出两个 worker 的任务说明里 Ownership 那一段。",
      "expected_output": "Each spec's Ownership picks 'Local commits on <short-branch> only, hooks must pass', keeps no push, merge into other branches, deploy or publish, and lists disjoint editable files per lane.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "Ownership 选“只在本 lane 短分支本地提交,hooks 必须通过”",
        "明确不 push、不合并到其他分支、不部署或发布",
        "两条 lane 的可编辑文件互不重叠"
      ]
    },
    {
      "id": 14,
      "name": "stop-mid-run-stops-what-it-can",
      "prompt": "这是演练,不要执行任何真实的 kill 或 orca 命令。你是总指挥,Run 台账在 fixtures/run-midrun(先复制到临时目录,只改副本)。B9 是 Orca worker(dispatch id ctx_b9),M3b 是你后台启动的进程(pid 4242),部署 test 由 CI 在跑、你没有权限取消。假设 worker-stop 和 kill 都会成功。我说:停,先别继续了。请逐条写出你会执行的停止命令(不执行),再按假设结果更新台账副本,最后回复我。",
      "expected_output": "Does not execute any real stop command; lists `orca orchestration worker-stop --dispatch ctx_b9` and `kill 4242` as the commands it would run; creates STOP in the copied Run; marks B9 and M3b stopped in the copy; leaves the deploy Wait running and explains it cannot be cancelled; WP6 is not started; tells the user in plain words what stopped and what still runs.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl",
        "fixtures/run-midrun/deploy-log.txt"
      ],
      "assertions": [
        "没有真的执行 kill 或 orca 停止命令",
        "列出会执行的 worker-stop(ctx_b9)和 kill 4242",
        "在副本里创建 STOP,WP6 不启动",
        "副本里 B9、M3b 标为 stopped,部署 test 仍为进行中并说明原因",
        "用白话告诉用户哪些已停、哪些还在跑,不让用户自己执行命令"
      ]
    },
    {
      "id": 15,
      "name": "reply-local-worker-question",
      "prompt": "回复 worker 问题:A1 按推荐。Worker 是协调者派出的本地会话。",
      "expected_output": "Read the Ask recommendation and linked session. Detect sideby send --help, send the exact answer when available, records the answer with ask answer, and reports the actual sideby receipt in chat without inventing success. Orca Worker uses orchestration reply. Use existing product fallback if unsupported. No per-message approval for this local Worker; external messages still need authorization.",
      "files": [
        "fixtures/run-sessions/run.json",
        "fixtures/run-sessions/sessions.json",
        "fixtures/sideby-stub/sideby"
      ],
      "assertions": [
        "Detects send support",
        "Reports the actual delivered/read/paste-needed receipt without inventing success",
        "Uses orca orchestration reply only for a Lane that has an Orca dispatch id; with orca_run null it neither invents an Orca path nor sends an Orca Worker through sideby",
        "Does not request per-message approval for authorized local Worker"
      ]
    },
    {
      "id": 16,
      "name": "fix-round-files-old-result-and-keeps-watching",
      "prompt": "这是演练,不要真的启动任何 Worker。你是总指挥,Run 台账在 fixtures/run-midrun(先复制到临时目录,再用 --run 指向副本)。B9 的审查发现了 2 个问题,要让 B9 原来的 Claude 会话续跑修掉。写出你会依次执行的台账命令和续跑命令,然后说明之后你怎么等它做完。",
      "expected_output": "Before resuming, raises B9's round with a new --output stream file (not the ended B9.jsonl), points the resumed run's stdout at that file, and keeps a `watch` running in the background so B9's end wakes the Coordinator instead of waiting for the user; no foreground sleep; the user runs nothing.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl"
      ],
      "assertions": [
        "续跑前执行 lane set --id B9 --round 2,并带上新的 --output 文件,不复用 B9.jsonl",
        "续跑命令的输出重定向到这个新文件",
        "用后台方式运行 dispatch_board.py watch 等 B9 结束,而不是等用户回复",
        "没有前台 sleep,也不让用户执行命令",
        "台账里 B9 新一轮的 output 路径和续跑命令实际写入的文件是同一个(相对路径按 Run 目录解析,和绝对路径指向同一文件也算)"
      ]
    },
    {
      "id": 17,
      "name": "shared-contract-lands-before-fanout",
      "prompt": "需求:packages/core/src/types/user.ts 的 User 类型加一个 role 字段;packages/api/src/routes/users.ts 按 role 过滤返回;packages/ui/src/components/UserBadge.tsx 按 role 显示角标。三处文件互不重叠,机器上 Claude、codex、grok 都有多个账号。帮我并行派活,给派活计划。",
      "expected_output": "Does not fan out three writers at once: the User.role field is a contract both api and ui rely on, so one agent lands it first (or the whole change stays one writer, L1); only after it lands may api and ui fan out. States the reason in the dispatch line.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "不同时派三个并行工人直接开写",
        "指出 User.role 是 api 和 ui 都依赖的共享契约,要先由一个 agent 定下并落地",
        "契约落地后才把 api 和 ui 分开并行,或说明为何整体由一个工人/总指挥做",
        "派活行里写出这个决定的理由"
      ]
    },
    {
      "id": 18,
      "name": "l1-coordinator-writes-itself",
      "prompt": "packages/api/src/middleware/rateLimit.ts 现在按 IP 限流,改成按登录用户限流,会动 rateLimit.ts、它的测试和 config.ts 三个文件。我机器上 codex、grok、三个 Claude 账号都能用。按你的派活方式安排。",
      "expected_output": "Picks L1 and the coordinator writes it itself: one module, three files, and nothing a Worker would buy (no crowded context, no other product's ability needed, quota available). Reviewer is another product. No Run Board, since no Worker is dispatched.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "选 L1(单模块的行为改动)",
        "总指挥自己写,不派写代码的工人",
        "安排一次独立评审,评审方是另一家产品或没写代码的新会话",
        "不为这件事开 Run 看板",
        "只跑受影响的测试"
      ]
    },
    {
      "id": 19,
      "name": "small-ux-tweak-is-not-research",
      "prompt": "设置页的「保存」按钮现在在底部,改到右上角,packages/web/src/pages/Settings.tsx 一个文件。按你的派活方式安排一下。",
      "expected_output": "Picks L0 (one file, layout only) and does it directly with a short recommendation; does not start the Research tier or send researchers, since the user did not ask how others do it and the answer is a taste call.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "选 L0 或 L1,总指挥直接改",
        "不进调研档,不派调研者",
        "不派工人、不开 worktree"
      ]
    },
    {
      "id": 20,
      "name": "worker-summary-carries-decisions",
      "prompt": "你是总指挥。给 L2 里 B 这条 lane(packages/core 的缓存层改造)写任务说明里的 Return 那一段,并说明 B 做完以后你会读哪些东西来决定怎么集成。",
      "expected_output": "The Return section asks for a summary of at most 10 lines: a first plain sentence for the user ending with a full stop, then lines for the Coordinator listing Decisions, Deviations, UNVERIFIED and Files (none when empty), plus a full report. The coordinator says it reads the whole summary, technical lines included, and not the full report or the jsonl stream.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "摘要不超过 10 行,第一句是给用户看的白话,以句号结尾,不含路径和 id",
        "摘要后面几行写给总指挥:做了哪些决定、哪里偏离了任务说明、哪些没验证(UNVERIFIED)、改了几个文件;空的写 none",
        "总指挥集成前读整份摘要(含技术行),不直接读完整报告或 jsonl 流"
      ]
    },
    {
      "id": 21,
      "name": "coordinator-writes-the-article",
      "prompt": "导出功能刚集成完,代码都是你派人写、你集成的。现在要写一篇给团队的技术分享讲它怎么实现的。机器上 agy、codex 都能用。按你的方式安排。",
      "expected_output": "Picks the Writing tier with the coordinator as writer, because it holds the implementation context; does not hand drafting to agy unless the user asks or its context is crowded; checks every fact, number, path and command against the code.",
      "files": ["fixtures/probe-full.json"],
      "assertions": [
        "由总指挥自己写,不默认交给 agy 等起草 agent",
        "说出理由:总指挥掌握实现过程的上下文",
        "逐条核对文中的事实、数字、路径和命令"
      ]
    },
    {
      "id": 22,
      "name": "close-reports-whether-dispatch-paid-off",
      "prompt": "这是演练,不要启动或停止任何进程。你是总指挥,Run 台账在 fixtures/run-midrun(先复制到临时目录,再用 --run 指向副本)。假设里面的任务都已经集成完了,帮我收尾,并告诉我这次派活值不值。",
      "expected_output": "Reads show for per-Lane cost and time, records a decision '派活是否值得' with the numbers (cost, time, fix rounds, integration problems) and a solo estimate labeled 估计, then closes the copy, or, when the ledger still has a running Wait or an open Ask, says what blocks closing instead of closing early. Does not claim a measured comparison with doing it solo.",
      "files": ["fixtures/run-midrun"],
      "assertions": [
        "先复制台账到临时目录,只操作副本",
        "用 show 的数据给出花费和用时(未知的不当成 0)",
        "用 decision add 记一条“派活是否值得”,写明数字",
        "“自己做会怎样”标成估计,不说成实测",
        "收尾:没有挡住的事项时 close;台账里还有进行中的外部等待或未答的问题时,说明什么挡着、不提前归档"
      ]
    },
    {
      "id": 23,
      "name": "unrelated-request-opens-a-side-session",
      "prompt": "这是演练,不要真的创建 worktree、终端或会话,只写出你会执行的命令和要交出去的任务说明。你是总指挥,正在给 packages/web 的章节编辑器做 B9 卡片改造,Run 台账在 fixtures/run-midrun(先复制到临时目录,再用 --run 指向副本),B9 的 Worker 还在写。这台机器上 Orca 在运行,我就在 Orca 里看着;sideby 也装了;目前没有你另开的会话。我突然想起来另一件事:scripts/release-notes.mjs 生成的更新日志把日期排反了(应该新的在前),帮我修一下。",
      "expected_output": "Treats the release-notes fix as unrelated to the B9 task (no shared files, no dependence on B9's unlanded result, its own acceptance check), so it neither adds a Lane to the current Run nor fixes it in this session, and does not ask the user whether to open a session. Writes a self-contained brief (the user's words, the change, an acceptance check, its own worktree, no push or deploy) to a file, then lists the commands it would run without running them: `orca worktree create --name side-<slug> --no-parent` and, in that worktree's own first terminal, `orca terminal send --terminal <handle> --enter --text \"<start> ...\"`, with no --focus or --activate. The account comes from the probe: this probe has no allow-list, so `<start>` may be `sideby next <family> --`, or `sideby run <ref> --` for an account the probe lists; an account outside the probe is wrong. The trace shows no `lane add` for this request and no orca or session-start command actually run. Tells the user in one sentence where the new session is and goes back to B9.",
      "files": [
        "fixtures/probe-full.json",
        "fixtures/run-midrun/run.json",
        "fixtures/run-midrun/B9.jsonl",
        "fixtures/run-midrun/deploy-log.txt"
      ],
      "assertions": [
        "判断这件事与当前任务不相关,并说出依据(不共享文件、不依赖 B9 未落地的结果、验收各自独立,说到其中至少两点)",
        "不往当前 Run 里加任务(Lane):执行过的命令里没有为这件事跑 lane add;也不在当前会话里直接修",
        "不先问用户要不要另开会话,直接安排",
        "写出一份自包含的任务说明:有用户原话或需求、验收方式、它自己的 worktree,不依赖当前会话的上下文",
        "在 Orca 里新开工作区,并在它自带的终端里启动会话(orca worktree create 加 orca terminal send),不带 --focus 或 --activate;账号取自探测结果:这份探测结果没有白名单,所以由 sideby next 选,或从探测列出的账号里选一个用 sideby run 启动",
        "用一句话告诉用户新会话开在哪,然后回到 B9",
        "执行过的命令里没有真的跑 orca worktree create、orca terminal send、orca agent launch 或启动会话的 sideby 命令;这些命令只出现在回复文字里"
      ]
    },
    {
      "id": 24,
      "name": "long-background-wait-gets-one-progress-line",
      "prompt": "这是演练,不要启动任何进程。你是总指挥,有一个 Run 开着:B9(写代码)25 分钟前派出,还在跑;部署 test 12 分钟前开始,还在构建;没有等我回答的问题。你 10 分钟前挂在后台的 watch 刚刚超时返回,没有任何任务结束;这 10 分钟里你没有给我发过任何文字,我也没有发消息。写出你现在发给我的原话,再说明之后你怎么继续等。",
      "expected_output": "Sends the user one unprompted line: what it is waiting on (B9 and the test deploy), how long so far, and what happens next. The line is not a milestone report: no per-Lane details, no log excerpts, no new question. Then the reply says how it keeps waiting: it will start the background watch again with a cap of about 10 minutes (`watch --timeout 600`); the reply proposes no foreground sleep and gives the user nothing to run. Graded on the final reply only: this case has no ledger fixture, so what the agent actually ran is not judged.",
      "files": [],
      "assertions": [
        "回复里给出现在就发给用户的进度原话,而不是说继续沉默等待或等用户来问",
        "发给用户的进度只有一行,包含三项:在等什么(B9、部署 test)、已等多久、下一步是什么",
        "这一行不展开任务细节,不写成里程碑报告(不逐项列明细、不贴日志、不新增问题)",
        "回复里说明之后会重新在后台挂 watch,并带约 10 分钟的超时(如 --timeout 600);回复里没有提出用前台 sleep 等待,也没有让用户执行命令"
      ]
    }
  ]
}
finishing-a-development-branchobra/superpowersUse when implementation is complete, all tests pass, and you need to decide how to integrate the workStars: 322kUpdated Claude CodeCodexVerifiedInstall
systematic-debuggingobra/superpowersUse when encountering any bug, test failure, or unexpected behavior, before proposing fixesStars: 322kUpdated Claude CodeCodexVerifiedInstall
test-driven-developmentobra/superpowersUse when implementing any feature or bugfix, before writing implementation codeStars: 322kUpdated Claude CodeCodexVerifiedInstall
skill-creatoranthropics/skillsCreate new skills, modify and improve existing skills, and measure skill performance. Use when users want to create a skill from scratch, edit, or optimize an existing skill, run evals to test a skill, benchmark skill performance with variance analysis, or optimize a skill's description for better triggering accuracy.Stars: 198kUpdated Claude CodeCodexVerifiedInstall