Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Claude DevFleet経由でマルチエージェントコーディングタスクをオーケストレーション — プロジェクトを計画し、分離された作業ツリー内で平行エージェントを派遣し、進捗を監視し、構造化レポートを読む。
.claude/skills/affaan-m-claude-devfleet/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 71% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 62% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 101% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 70% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 60% | 0% |
当需要调度多个 Claude Code 智能体并行处理编码任务时使用此技能。每个智能体在独立的 git worktree 中运行,并配备全套工具。
需要连接一个通过 MCP 运行的 Claude DevFleet 实例:
bashclaude mcp add devfleet --transport http http://localhost:18801/mcp
用户 → "构建一个带有身份验证和测试的 REST API"
↓
plan_project(prompt) → 项目ID + 任务DAG
↓
向用户展示计划 → 获取批准
↓
dispatch_mission(M1) → 代理1在工作树中生成
↓
M1完成 → 自动合并 → 自动分发M2 (依赖于M1)
↓
M2完成 → 自动合并
↓
get_report(M2) → 更改的文件、完成的工作、错误、后续步骤
↓
向用户报告| 工具 | 用途 | |------|---------| | plan_project(prompt) | AI 将描述分解为包含链式任务的项目 | | create_project(name, path?, description?) | 手动创建项目,返回 project_id | | create_mission(project_id, title, prompt, depends_on?, auto_dispatch?) | 添加任务。depends_on 是任务 ID 字符串列表(例如 ["abc-123"])。设置 auto_dispatch=true 可在依赖满足时自动启动。 | | dispatch_mission(mission_id, model?, max_turns?) | 启动智能体执行任务 | | cancel_mission(mission_id) | 停止正在运行的智能体 | | wait_for_mission(mission_id, timeout_seconds?) | 阻塞直到任务完成(见下方说明) | | get_mission_status(mission_id) | 检查任务进度而不阻塞 | | get_report(mission_id) | 读取结构化报告(更改的文件、测试情况、错误、后续步骤) | | get_dashboard() | 系统概览:运行中的智能体、统计信息、近期活动 | | list_projects() | 浏览所有项目 | | list_missions(project_id, status?) | 列出项目中的任务 |
> 关于 wait_for_mission 的说明: 此操作会阻塞对话,最长 timeout_seconds 秒(默认 600 秒)。对于长时间运行的任务,建议改为每 30-60 秒使用 get_mission_status 轮询,以便用户能看到进度更新。
plan_project(prompt="...") → 返回 project_id 以及带有 depends_on 链和 auto_dispatch=true 的任务列表。depends_on 为空)调用 dispatch_mission(mission_id=<first_mission_id>)。剩余任务在其依赖项完成时自动调度(因为 plan_project 为它们设置了 auto_dispatch=true)。get_mission_status(mission_id=...) 或 get_dashboard() 检查进度。get_report(mission_id=...)。与用户分享亮点。DevFleet 默认最多同时运行 3 个智能体(可通过 DEVFLEET_MAX_AGENTS 配置)。当所有槽位都占满时,设置了 auto_dispatch=true 的任务会在任务监视器中排队,并在槽位空闲时自动调度。检查 get_dashboard() 了解当前槽位使用情况。
plan_project(prompt="...") → 显示包含任务和依赖关系的计划。depends_on 为空的那个)。auto_dispatch=true)。get_mission_status 或 get_dashboard() 轮询,直到所有任务达到终止状态(completed、failed 或 cancelled)。get_report(mission_id=...)——总结成功之处,并指出失败任务及其错误和后续步骤。create_project(name="My Project") → 返回 project_id。create_mission(project_id=project_id, title="...", prompt="...", auto_dispatch=true) → 捕获 root_mission_id。为每个后续任务执行 create_mission(project_id=project_id, title="...", prompt="...", auto_dispatch=true, depends_on=["<root_mission_id>"])。
dispatch_mission(mission_id=...) 以启动链。get_report(mission_id=...)。create_project(name="...") → 获取 project_id。create_mission(project_id=project_id, title="Implement feature", prompt="...") → 获取 impl_mission_id。dispatch_mission(mission_id=impl_mission_id),然后使用 get_mission_status 轮询直到完成。get_report(mission_id=impl_mission_id) 以审查结果。create_mission(project_id=project_id, title="Review", prompt="...", depends_on=[impl_mission_id], auto_dispatch=true) —— 由于依赖已满足,自动启动。get_dashboard() 了解智能体槽位可用性。auto_dispatch=true。没有此标志,任务将保持 draft 状态。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-09 | fail→fail | 7,559 | 2,743 | -64% | 1 | 1 | 0% | 1,294 | 2,091 | +62% | 0 | 0 | — |
case-01 | fail→fail | 24,456 | 6,194 | -75% | 1 | 1 | 0% | 6,209 | 2,010 | -68% | 0 | 0 | — |
case-02 | fail→fail | 2,298 | 6,313 | +175% | 1 | 1 | 0% | 363 | 1,930 | +432% | 0 | 0 | — |
case-03 | fail→fail | 2,371 | 3,385 | +43% | 1 | 1 | 0% | 421 | 1,841 | +337% | 0 | 0 | — |
case-22 | pass→pass | 8,796 | 5,782 | -34% | 1 | 1 | 0% | 1,757 | 2,700 | +54% | 0 | 0 | — |
case-04 | fail→pass | 6,605 | 3,249 | -51% | 1 | 1 | 0% | 1,349 | 2,303 | +71% | 0 | 0 | — |
case-05 | fail→pass | 8,227 | 4,938 | -40% | 1 | 1 | 0% | 1,576 | 2,555 | +62% | 0 | 0 | — |
case-06 | fail→pass | 5,850 | 2,252 | -62% | 1 | 1 | 0% | 983 | 1,973 | +101% | 0 | 0 | — |
case-07 | fail→pass | 6,379 | 7,067 | +11% | 1 | 1 | 0% | 1,068 | 1,813 | +70% | 0 | 0 | — |
case-08 | fail→pass | 7,845 | 3,666 | -53% | 1 | 1 | 0% | 1,478 | 2,365 | +60% | 0 | 0 | — |
case-10 | pass→fail | 8,050 | 2,407 | -70% | 1 | 1 | 0% | 1,481 | 1,947 | +31% | 0 | 0 | — |
case-11 | pass→pass | 8,315 | 2,580 | -69% | 1 | 1 | 0% | 1,376 | 2,009 | +46% | 0 | 0 | — |
case-12 | pass→pass | 6,894 | 3,189 | -54% | 1 | 1 | 0% | 1,000 | 2,051 | +105% | 0 | 0 | — |
case-13 | fail→pass | 8,340 | 1,587 | -81% | 1 | 1 | 0% | 1,544 | 1,844 | +19% | 0 | 0 | — |
case-14 | fail→pass | 7,810 | 1,810 | -77% | 1 | 1 | 0% | 1,284 | 1,848 | +44% | 0 | 0 | — |
case-15 | pass→pass | 4,420 | 3,862 | -13% | 1 | 1 | 0% | 781 | 2,212 | +183% | 0 | 0 | — |
case-16 | pass→pass | 6,542 | 3,011 | -54% | 1 | 1 | 0% | 1,077 | 2,077 | +93% | 0 | 0 | — |
case-17 | fail→pass | 9,510 | 3,734 | -61% | 1 | 1 | 0% | 1,640 | 2,134 | +30% | 0 | 0 | — |
case-18 | pass→pass | 4,136 | 1,521 | -63% | 1 | 1 | 0% | 712 | 1,800 | +153% | 0 | 0 | — |
case-19 | fail→fail | 8,565 | 2,795 | -67% | 1 | 1 | 0% | 1,696 | 2,020 | +19% | 0 | 0 | — |
case-20 | fail→pass | 9,719 | 3,281 | -66% | 1 | 1 | 0% | 1,822 | 2,169 | +19% | 0 | 0 | — |
case-21 | pass→pass | 2,825 | 2,051 | -27% | 1 | 1 | 0% | 497 | 1,914 | +285% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +36 percentage points is the difference between those two pass rates over the 20 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.