OMO使用介绍

关于本笔记

本笔记用于沉淀 oh-my-opencode(OMO)插件在 OpenCode 中的使用方法与适用场景,目标是理清四类 Agent 的职责边界,以及如何与 Superpowers skills、vibe 治理流程配合,形成可重复的开发工作流。

项目地址:https://github.com/code-yeongyu/oh-my-openagent

目录

项目简介

oh-my-opencode(OMO)是一套面向软件开发的 Agent 编排系统,它不是单一聊天机器人,而是把开发过程拆成不同角色:

  • 规划 — 想清楚怎么做
  • 深度执行 — 深入解决问题
  • 编排 — 统筹全局推进
  • 按计划落地 — 照既定计划执行

一句话:用对的人做对的事,不把所有期望压在一个 Agent 上。

安装配置参考https://www.opencodecn.com/docs/best-practices/oh-my-opencode

项目结构概览(四类 Agent)

OMO 通过 oh-my-opencode 插件注册到 OpenCode,提供四类预定义 Agent:

1
2
3
4
5
6
7
8
9
10
11
12
oh-my-opencode(OMO 插件)
├─ Prometheus(规划者)
│ 把模糊需求拆成可执行计划,保存为 .sisyphus/plans/*.md

├─ Hephaestus(深度执行者)
│ 端到端解决问题:探索 → 定位 → 实现 → 验证 → 修复

├─ Sisyphus(编排器/总控)
│ 统筹复杂任务:决定何时规划、搜索、查文档、派子 Agent、整体验证

└─ Atlas(计划执行者)
读取已有计划,按步骤推进,不重新设计

配合的外部 Agent 角色:

  • explore — 代码库快速搜索(contextual grep)
  • librarian — 外部文档/官方 API 查询
  • oracle — 复杂架构咨询与调试
  • Momus — 计划/工作审查

四类 Agent 的职责

Prometheus:Plan Builder,规划者

Prometheus 负责把目标拆成可执行计划。它适合需求不够清晰、影响范围较大、需要先设计方案时使用。

典型流程:

  1. 理解目标、约束和验收标准。
  2. 分析代码结构、依赖关系和风险点。
  3. 拆分阶段、任务、文件范围和验证方式。
  4. 输出可执行计划,形式是 .sisyphus/plans/*.md
  5. 后续交给 Atlas、Hephaestus 或 Sisyphus 执行。

适合的请求:

1
先用 Prometheus 给这个重构做实施计划,不要改代码。
1
这个功能影响范围比较大,先帮我拆解方案和风险。

Hephaestus:Deep Agent,深度执行者

Hephaestus 负责端到端解决一个明确目标。它会探索代码、定位问题、实现修改、运行验证,并根据验证结果继续修复。

典型流程:

  1. 判断用户真实意图:分析、修复、实现、重构还是调研。
  2. 阅读代码库上下文,不凭空猜测。
  3. 必要时调用 explore、librarian、oracle 等辅助 Agent。
  4. 修改代码或文档。
  5. 运行诊断、测试、类型检查或构建。
  6. 修复验证阶段发现的问题。
  7. 输出完成内容和验证证据。

适合的请求:

1
使用 Hephaestus 深入排查这个性能问题,找到根因并修复。
1
帮我修复登录页 token 过期后无法刷新的问题,并跑相关测试。

Sisyphus:Ultraworker,编排器 / 总控

Sisyphus 负责统筹复杂任务。它像项目总控,决定何时规划、何时搜索、何时查文档、何时分派子 Agent,以及何时做整体验证。

典型流程:

  1. 接收目标并判断复杂度。
  2. 需要时调用 Prometheus 生成计划。
  3. 将不同子任务分派给 Hephaestus、Atlas、explore、librarian 等。
  4. 收集子任务结果。
  5. 处理冲突、依赖和验证缺口。
  6. 推动任务直到收敛。

适合的请求:

1
使用 Sisyphus 统筹完成这个 issue:先规划,再实现,再验证。
1
这个需求涉及多个模块,请自动拆分、执行并验证。

Atlas:Plan Executor,计划执行者

Atlas 负责执行已经确认的计划。它不适合重新设计方案,而适合按步骤稳定推进。

典型流程:

  1. 读取已有计划。
  2. 按计划步骤执行。
  3. 每完成一步做局部验证。
  4. 遇到计划不清楚或执行失败时反馈。
  5. 最后运行整体验证。

适合的请求:

1
使用 Atlas 执行 .sisyphus/plans/payment-refactor.md,不要重新规划。
1
这个计划已经确认了,请按步骤实现并验证。

Agent 组合关系与流转

大型任务可以按如下方式流转:

1
2
3
4
5
6
7
8
9
10
11
12
13
用户目标

Sisyphus 判断复杂度并统筹

Prometheus 生成计划

人工或 Momus 审查计划,可选

Atlas 按计划执行

Hephaestus 处理复杂子任务或深水区问题

Sisyphus 汇总、验证、收尾

轻量任务不需要完整链路。例如一个小 bug 可以直接交给 Hephaestus;一个已有计划可以直接交给 Atlas。

分类使用场景

日常小修小改

适合 bug fix、小功能、测试失败、类型错误、局部重构。

1
修复这个报错,并跑相关测试。
1
给这个组件补一个 loading 状态,保持现有样式风格。

不确定怎么做

适合影响多个模块、业务规则复杂、技术方案不明确、需要先审查方案的任务。

1
先用 Prometheus 规划这个权限系统改造,不要修改代码。

已有计划

适合已经确认实施方案,不希望 Agent 重新设计的任务。

1
使用 Atlas 执行 .sisyphus/plans/auth-refactor.md,按步骤验证。

复杂端到端任务

适合多模块、多阶段、需要搜索代码、查文档、实现和验证的任务。

1
使用 Sisyphus 统筹完成这个需求:先规划,再实现,再验证。

难 bug 或深水区问题

适合性能问题、复杂业务逻辑、类型系统问题、难以复现的 bug。

1
使用 Hephaestus 深入排查这个问题,定位根因、修复并验证。

选择 Agent 速查表

你的需求 推荐用法
小修小改 直接描述目标,让 Hephaestus 执行
不确定怎么做 先让 Prometheus 出计划
已有明确计划 让 Atlas 按计划执行
大型端到端任务 让 Sisyphus 统筹
难查的 bug / 性能问题 让 Hephaestus 深入排查
只想了解代码 明确说”只分析,不修改文件”
涉及外部库或框架 让流程加入 librarian / documentation lookup
需要浏览器验证 让流程加入 playwright

推荐提示词模板

只规划,不改代码

1
使用 Prometheus 规划这个任务。只输出计划、影响范围、风险和验证方式,不修改文件。

按计划执行

1
使用 Atlas 执行 .sisyphus/plans/xxx.md。不要重新规划,按步骤实现并验证。

端到端完成

1
使用 Sisyphus 统筹完成这个需求:先理解目标,再规划,实现后运行相关测试和构建。

深入排查并修复

1
使用 Hephaestus 深入排查这个 bug,找到根因,修复后运行相关验证。

明确禁止修改

1
只分析,不修改任何文件。请说明当前实现、问题点和建议方案。

使用 OMO 默认编排

1
本次按 OMO 默认方式执行,不启用 vibe 流程。

使用 vibe 治理

1
本次按 vibe 流程执行,禁用自动花式分派。

并发与安全边界

OMO 支持并发子代理,但是否并发取决于调用方式和任务类型。

常见并发场景:

  • 多个 explore 同时搜索不同代码区域。
  • librarian 查询外部文档时,主 Agent 做本地非重叠探索。
  • 多个只读分析任务并行执行。

需要谨慎的场景:

  • 多个 Agent 同时修改同一文件。
  • 多个 Agent 同时重构同一个模块。
  • 一个任务依赖另一个任务的结果,却被提前并发执行。

经验原则:

探索可以并发,写代码要控制边界;只读任务适合并行,写入任务需要明确文件范围。

OMO 与 vibe 的关系

一次任务只选一个主控框架:

  • 如果选择 OMO 主导,就让 Sisyphus / Hephaestus / Prometheus / Atlas 按 OMO 方式推进。
  • 如果选择 vibe 主导,就明确说”本次按 vibe 流程执行”。
  • 不建议在同一个任务中同时叠加 OMO 自由编排和 vibe 治理流程,避免两套调度逻辑互相覆盖。

vibe 是治理型运行协议,适合中大型任务的约束、计划和阶段清理。OMO 是 Agent 编排系统,适合自动分派、深度执行和上下文探索。

推荐做法:

  • 日常快速开发:OMO 主导。
  • 大型、风险高、需求需要冻结的任务:vibe 主导。
  • 如果 vibe 主导,OMO/skills 可以作为执行阶段的能力补充,但不再另起一套自由编排。

项目级增强:AGENTS.md 与任务技能映射

默认 OMO 的 Agent 编排已经能完成很多工作,但在项目级别还可以进一步增强。核心思路是:不要只依赖 OMO 的默认 Agent 行为,而是在项目里写清楚 Agent 应该如何选择 skills、如何与 Superpowers 工作流配合、何时切换到 vibe 治理流程。

项目通过两个文件做增强:

  • AGENTS.md:告诉所有 Agent 本项目的协作原则、运行时边界、可用 skills、OMO 与 vibe 的关系。
  • .sisyphus/task-skill-map.md:告诉 Sisyphus 在分派子任务时,应该如何把任务类型映射到 load_skills

为什么需要项目级定制

默认 OMO 可能知道有哪些 Agent,但不一定天然知道你的项目希望如何使用 skills。例如:

  • Bug 修复是否必须先诊断?
  • 前端视觉任务是否必须加载 frontend-designfrontend-ui-ux
  • 实现任务是否默认走 TDD?
  • 完成前是否必须运行验证?
  • 大型任务是交给 OMO 统筹,还是交给 vibe 治理?

这些偏好不应该每次都靠口头提醒,而应该沉淀到项目级文档中。

AGENTS.md 核心原则

AGENTS.md 适合放高层规则,例如:

  • 先理解问题,再选择 skill。
  • 证据优先,不能在没有验证时声称完成。
  • 一次任务只选一个主控框架:OMO 或 vibe。
  • 轻量任务走最短路径,中大型任务可以进入治理流程。
  • 子 Agent 分派时不能默认 load_skills=[]

任务技能映射表(.sisyphus/task-skill-map.md)

Sisyphus 通过 task() 分派子 Agent 时,必须把任务类型映射为明确的 load_skills,不要把实现、调试、重构、审查任务默认派成 load_skills=[]

选择技能时区分两类:

  • 工作流技能:负责”怎么推进” — brainstormingwriting-planstest-driven-developmentsubagent-driven-developmentverification-before-completion
  • 领域/实现技能:负责”怎么写对” — context-hunteraios-devfrontend-designdiagnose

常用基线:

任务类型 推荐 category 推荐 load_skills
后端/API/认证/持久层 deepquick context-hunter, aios-dev, tdd
全栈功能 deep context-hunter, autonomous-builder, aios-dev, tdd
前端 UI/布局/交互 visual-engineering context-hunter, frontend-design, frontend-ui-ux, tdd
Bug/失败测试/异常行为 deepquick diagnose, context-hunter, tdd
架构/重构/模块边界 ultrabraindeep context-hunter, zoom-out, improve-codebase-architecture
代码审查 unspecified-high code-reviewer, reviewing-code, code-review-excellence
Git 操作 quickunspecified-high git-master

Superpowers 与 OMO 的配合方式

Superpowers 是”工作方法论 skills”,OMO 是”Agent 编排系统”。两者配合但不竞争。

常见组合:

  • 需求不清楚:brainstorming + Prometheus
  • 计划编写:writing-plans + Prometheus
  • 按计划执行:executing-planssubagent-driven-development + Atlas
  • 实现/修复:test-driven-developmenttdd + Hephaestus
  • 完成前验证:verification-before-completion + Sisyphus/Hephaestus
  • 开发分支收尾:finishing-a-development-branch + Git

关键原则:OMO 决定”谁来做”,Superpowers 约束”怎么做”。

推荐的项目级配置结构

1
2
3
4
5
AGENTS.md                        # 全局协作规则
.sisyphus/
task-skill-map.md # 任务到 skills 的映射
plans/
<task-plan>.md # Prometheus 生成的执行计划

配置文件说明(重点)

使用 OMO 后重点关注以下配置文件:

  • opencode.json
    • 插件加载入口,注册 oh-my-openagent 和 superpowers
  • AGENTS.md
    • 项目级 AI 协作规则、Skill 选择原则、Agent 职责边界
  • .sisyphus/task-skill-map.md
    • Sisyphus 分派子 Agent 时 load_skills 的映射表,避免默认传空
  • .sisyphus/plans/*.md
    • Prometheus 生成的执行计划

常见误区与修正

误区 1:让 Sisyphus 做所有事,不区分 Agent 角色。

修正:Sisyphus 是编排器,不是执行器。实现交给 Hephaestus,计划交给 Prometheus,已有计划交给 Atlas。

误区 2:子 Agent 分派时默认 load_skills=[]

修正:实现、调试、重构、审查任务都应带上明确 skills。只有 explore/librarian 等纯搜索 agent 才适合空列表。

误区 3:同时启用 OMO 自由编排和 vibe 治理流程。

修正:一次任务只选一个主控框架。OMO 主导就禁用 vibe,vibe 主导就禁用 OMO 自由编排。

误区 4:未澄清需求直接让 Agent 写代码。

修正:大任务先用 Prometheus 规划,难问题先用 Hephaestus 排查,边界模糊时先用 brainstorming。

误区 5:完成声明没有验证证据。

修正:任何”完成/修复/通过”声明,必须附验证命令和输出。verification-before-completion skill 强制此规则。

我的最小实践模板

以下的AGENTS.md与task-skill-map.md不一定符合你的项目, 使用AI根据你的实际情况进行重构

AGENTS.md:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
# AI 协作启示(本地 Skills 能力地图)

> 目的:这是一个**启示型**协作文档,不是死板流程表。
> 目标:让 AI 知道你本地具备大量 Skills,可按任务实际情况动态组合调用。

## 核心原则

1. **先理解问题,再选 Skill**:先澄清目标、约束、验收标准,再决定调用路径。
2. **组合是动态的,不设上限**:Skill 可以多段拼接,按任务复杂度增减,不固定套餐。
3. **证据优先**:任何“完成/修复/通过”声明,必须附验证证据(命令、输出、检查点)。
4. **单一运行时权威**:涉及实现任务时,优先沿 `vibe` 治理链路推进,避免并行造第二套路由。
5. **最小充分调用**:优先少而准;当任务复杂时,再扩展到多 Skill/多 Agent。

## 本地 Skills 现状(能力提示)

当前环境已安装大规模 Skills,覆盖但不限于:

- **工程开发**:调试、测试、评审、CI、部署、性能、架构
- **数据与 ML**:EDA、统计建模、可视化、训练与评估
- **文档与知识**:Markdown/Docx/PDF/PPTX、报告、引用、知识沉淀
- **科研生信与化学**:多数据库查询、组学分析、药化与结构相关流程
- **平台集成**:GitHub、OpenAI、Sentry、Vercel、Netlify、MCP 等

> 注意:以上只是能力地图,不是限制列表;未列出的技能同样可以被调用。

## 智能体情况

除了vibe的智能体(Vibe-Review, Vibe-Implement, Vibe-Plan), 还有来自 oh-my-opencode 插件(oh-my-opencode@latest注入到 OpenCode 的“agent 注册表”)的智能体

- Prometheus (Plan Builder):偏“规划/拆解”。把目标拆成可执行步骤/计划(文档里称它是 planner),通常配合 Sisyphus 的编排工作流使用。
- Hephaestus (Deep Agent):偏“深度执行”。更像 autonomous worker,适合端到端实现一个较大的子任务(需要自己探索、自己完成)。
- Sisyphus (Ultraworker):编排器/总控。负责决定何时探索、何时查文档、何时把任务拆给子 agent、以及验证/推进整体流程。
- Atlas (Plan Executor):偏“按计划执行”。把已经写好的计划按步骤推进执行(相当于执行器/协调执行)。

## 推荐工作方式(非强制)

### 轻量任务

- 直接执行最短路径(可不进入复杂编排)
- 只在必要时调用 1-2 个针对性 Skills

### 中大型任务

- 使用 `vibe` 的治理阶段进行约束与计划:
- `skeleton_check`
- `deep_interview`
- `requirement_doc`
- `xl_plan`
- `plan_execute`
- `phase_cleanup`
- 在执行阶段动态选择最合适 Skills(可并行)
- 搭配 Superpowers 插件提供的工作流技能,但不要让它成为第二套运行时:
- `brainstorming`:理解需求、澄清模糊点、形成设计。
- `writing-plans`:把设计拆成可执行计划。
- `test-driven-development` / `tdd`:实现与修复时执行红-绿-重构。
- `subagent-driven-development`:按计划分派子 Agent,并在任务间做审查。
- `requesting-code-review` / `reviewing-code` / `code-reviewer`:完成任务后审查。
- `verification-before-completion`:任何完成声明前先跑验证命令。
- `finishing-a-development-branch`:实现完成后的合并、PR、保留或清理选择。

### vibe or oh-my-opencode

- oh-my-opencode 是作为 OpenCode 插件加载的(C:\Users\<Users>\.config\opencode\opencode.json:28)。
- vibe 是作为 skill 体系中的运行协议(governed runtime)存在,不是另一个插件内核。

实操建议(最稳):

1. 一次任务只选一个主控框架:要么“vibe主导”,要么“oh-my-opencode主导”。
2. 需要 vibe 时,明确写:“本次按 vibe 流程执行,禁用自动花式分派”。
3. 日常快速开发就用 oh-my-opencode 默认自动编排,不额外叠加 vibe 管控。

### 研究/探索类任务

- 先做上下文收集(代码、数据、文档)
- 必要时引入专长 Skill(统计、可视化、数据库、文献)
- 输出结论时区分:已证实 / 待验证 / 假设

## 动态选型提示(给 AI)

当用户提出任务时,请按下列顺序自检:

1. 任务类型:实现 / 修复 / 分析 / 文档 / 调研 / 部署
2. 风险等级:是否涉及生产、数据安全、凭据、不可逆操作
3. 证据需求:需要哪些验证命令与验收标准
4. Skill 组合:从本地已装技能中选最匹配组合,必要时迭代替换

### Sisyphus 子 Agent 技能分发规则

Sisyphus 通过 `task()` 分派子 Agent 时,必须把任务类型映射为明确的 `load_skills`,不要把实现、调试、重构、审查任务默认派成 `load_skills=[]`

项目级映射表位于 `.sisyphus/task-skill-map.md`。分派前先按该表选择技能;若没有精确匹配,默认传入 `context-hunter`,涉及生产行为变更时再加入 `tdd`

选择技能时区分两类:Superpowers 工作流技能负责“怎么推进”(如 `brainstorming``writing-plans``test-driven-development``subagent-driven-development``verification-before-completion`),领域/实现技能负责“怎么写对”(如 `context-hunter``aios-dev``frontend-design``diagnose`)。大型任务可以两类组合;轻量任务优先选择最少必要技能。

允许 `load_skills=[]` 的主要场景是纯搜索型子 Agent,例如 `explore` / `librarian`,或者明确不需要技能的轻量文本整理。除此之外,空列表必须视为需要重新评估的异常。

常用基线:

- 后端/API/认证/持久层:`context-hunter`, `aios-dev`, `tdd`
- 前端/UI/布局/交互:`context-hunter`, `frontend-design`, `frontend-ui-ux`, `tdd`
- Bug/失败测试/异常行为:`diagnose`, `context-hunter`, `tdd`
- 架构/重构/模块边界:`context-hunter`, `zoom-out`, `improve-codebase-architecture`
- 代码审查:`code-reviewer`, `reviewing-code`, `code-review-excellence`
- Git 操作:`git-master`

子 Agent prompt 中也应显式写明:`Required skills for this task: ... Load and follow them before editing files.` 这只是兜底;首选机制仍然是 `task(load_skills=[...])`

## 禁止与边界

- 不得把“示例组合”当作唯一固定流程。
- 不得在证据不足时声称任务完成。
- 不得因某个 Skill 不匹配而停止思考;应改用更合适技能或回退基础链路。

## 一句话执行约定

> **以任务事实为中心,动态编排本地 Skills;先约束、后实现、再验证。**

task-skill-map.md:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80

# Sisyphus Task Skill Map

This file is the project-local source of truth for choosing `load_skills` when Sisyphus delegates work through `task()`.

## Core rule

Do not default implementation, review, debugging, or refactoring subagents to `load_skills=[]`.

Use an empty skill list only for pure discovery agents whose job is contextual search, such as `explore` and `librarian`, or when a task is explicitly skill-neutral.

## Default fallback

If no row clearly matches, use:

```text
load_skills=["context-hunter"]
```

If the task writes or changes production behavior, add `tdd` unless the task is explicitly documentation-only, generated output, or configuration-only.

## Skill families

Use two families together when the task is large enough:

- **Workflow skills** describe how to run the work: `brainstorming`, `writing-plans`, `test-driven-development`, `subagent-driven-development`, `executing-plans`, `requesting-code-review`, `verification-before-completion`, `finishing-a-development-branch`.
- **Domain skills** describe how to do the work well: `context-hunter`, `aios-dev`, `frontend-design`, `diagnose`, `documentation-lookup`, `improve-codebase-architecture`, file-format skills, and similar specialist skills.

Do not add every workflow skill to every subagent. Pick the stage-specific one: planning gets planning skills, implementation gets TDD/context skills, completion gets verification/finish skills.

## Task mapping

| Task shape | Recommended category | Required `load_skills` | Notes |
|---|---:|---|---|
| Backend API, service logic, persistence, auth, job processing | `deep` or `quick` by scope | `context-hunter`, `aios-dev`, `tdd` | Use `deep` for cross-module changes; `quick` only for narrow edits. |
| Full-stack feature touching frontend and backend | `deep` | `context-hunter`, `autonomous-builder`, `aios-dev`, `tdd` | Add frontend skills if visual/UI behavior changes. |
| Frontend UI, layout, CSS, component polish, UX flow | `visual-engineering` | `context-hunter`, `frontend-design`, `frontend-ui-ux`, `tdd` | Visual work must use `visual-engineering`; add `playwright` when browser verification is needed. |
| Browser automation, UI-flow verification, screenshots | `visual-engineering` or `quick` by scope | `playwright`, `playwright-cli` | Use with implementation skills when validating UI changes. |
| Bug, failing test, unexpected behavior, regression | `deep` or `quick` by scope | `diagnose`, `context-hunter`, `tdd` | Find root cause before fixing; prove with a failing test when possible. |
| Architecture review, refactor opportunity, module deepening | `ultrabrain` or `deep` | `context-hunter`, `zoom-out`, `improve-codebase-architecture` | Add `tdd` when code changes will follow. |
| Code review, PR review, quality check | `unspecified-high` | `code-reviewer`, `reviewing-code`, `code-review-excellence` | For post-implementation review, prefer `review-work` when available. |
| Documentation or prose-only changes | `writing` | `writing-plans` when planning, otherwise no required implementation skill | Do not force `tdd` for prose-only work. |
| Requirements clarification, design exploration, product tradeoffs | `deep` or `writing` | `brainstorming` | Use before creative behavior/design work; stop before implementation until design is approved. |
| Implementation plan creation from an approved design/spec | `deep` or `writing` | `writing-plans` | Save plans under the project's chosen plan location; include exact files, tests, and commands. |
| Saved plan execution with mostly independent tasks | `deep` | `subagent-driven-development`, `context-hunter`, `tdd`, `requesting-code-review` | If executing a saved plan, pass the relevant per-task domain skills too. |
| Saved plan execution in a single session without subagent split | `deep` | `executing-plans`, `verification-before-completion` | Prefer `subagent-driven-development` when subagents are available. |
| Before claiming completion/fixed/passing | same as parent task | `verification-before-completion` | Run fresh verification before any completion claim. |
| Implementation complete and user asks how to finish/merge/PR | `quick` or `unspecified-high` | `finishing-a-development-branch`, `git-master` | Verify tests before offering merge/PR/keep/discard options. |
| Vibe-governed implementation | `deep` | `vibe`, plus the row-specific skills above | `vibe` owns runtime stages; Superpowers skills provide stage discipline only. |
| Git commit, branch, PR, blame, rebase, squash | `quick` or `unspecified-high` by risk | `git-master` | Never commit unless the user explicitly asks. |
| File write failure or high-risk file generation | same as parent task | `smart-file-writer` | Use reactively on write errors or proactively for critical writes. |
| Library/framework/API usage needing current docs | same as parent task | `documentation-lookup` | Also fire `librarian` for external examples or official docs research. |
| Word/PDF/PPTX/XLSX work | `writing` or domain category | `docx`/`pdf`/`pptx`/`spreadsheet`/`xlsx` as applicable | Use the file-format-specific skill before editing or generating artifacts. |

## Delegation prompt contract

When delegating implementation work, include a short line near the top of the subagent prompt:

```text
Required skills for this task: <comma-separated skill names from this map>. Load and follow them before editing files.
```

This text is a backup guard. The primary mechanism is still the `task(load_skills=[...])` argument.

## Examples

```text
Backend bug fix:
task(category="deep", load_skills=["diagnose", "context-hunter", "tdd"], ...)
```
```text
React page polish:
task(category="visual-engineering", load_skills=["context-hunter", "frontend-design", "frontend-ui-ux", "tdd", "playwright"], ...)
```

```text
Architecture refactor candidate:
task(category="ultrabrain", load_skills=["context-hunter", "zoom-out", "improve-codebase-architecture"], ...)
```