od

name	od
description	OmniDev AI-driven development workflow. Use ONLY when the user's message starts with /od (e.g. /od h, /od re, /od ob, /od rp, /od rv, /od qa, /od ch, /od ln). Do not load or follow this skill for normal chat without the /od prefix.

OmniDev Workflow Skill — Full Specification

This file is the single source of truth for all OmniDev rules. The lightweight trigger lives in rules/01-omnidev-workflow.mdc (alwaysApply: false); everything below applies only when the current user message starts with /od — including workflow, docs/omnidev-state/**, and evolution logging. No OmniDev behavior on non-/od turns.

---

A. Command Reference

All commands support short aliases (1-2 letters). Users can also reply with numbers (e.g. 1, 2, 3) when presented with numbered options at any checkpoint.

Core Commands

Command	Alias	说明
/od [需求]	—	引导式工作流：评估复杂度 → 推荐阶段 → 可跳过任意阶段
/od -f [需求]	—	快速模式：跳过蓝图/计划，直接开发（热修复）
/od -p [需求]	—	仅规划：只输出蓝图和计划，不写代码
/od h	/od help	显示所有命令
/od ob	/od onboard	扫描项目，生成上下文文档
/od rp	/od report	生成周报
/od rv	/od review	代码审查（只读，不修改）
/od qa	—	依赖分析 → Mock → 场景覆盖 → 韧性测试 → 测试报告
/od ch [新需求]	/od change	需求变更管理
/od ln	/od learn	自学习：回顾错误 + 提炼规则 + 演化提案
/od ln -r	—	查看学习日志和待处理提案
/od ln -a	—	自动应用所有待处理提案
/od ln --rb [N]	—	回滚第 N 条演化
/od up	/od update	更新 OmniDev Kit 到最新版本
/od i <url>	/od install	从远程 Git 仓库安装 OmniDev Kit
/od ps	/od push	提交并推送代码
/od st	/od stash	暂存当前任务上下文
/od po	/od pop	恢复暂存的任务上下文
/od sy	/od sync	同步输出到 Jira/GitHub Issue
/od db	/od dashboard	生成全局效率 ROI 面板
/od re	/od resume	恢复上次中断的会话（加载 OmniDev 规则）
/od cfg	/od config	查看当前 OmniDev 配置
/od cfg -i on	—	开启交互模式 + 自动问答模式（决策点使用结构化选择 UI；任务结束后自动切换 Ask 模式；默认开启）
/od cfg -i off	—	关闭交互模式 + 自动问答模式

Phase Navigation (阶段导航)

Command	Alias	说明
/od n	/od next	下一阶段
/od ad [内容]	/od adj	修订当前阶段输出
/od sk [阶段]	/od skip	跳过某个阶段
/od bk [阶段]	/od back	返回某个阶段
/od al	/od all	执行所有剩余阶段（不暂停）

Confirmation (确认操作)

At every checkpoint, options are presented as numbered list — user can reply with the number, the alias, or the full command. Example: reply 1 or /od n or /od next all mean "proceed to next phase".

Command	Alias	说明
/od y	/od confirm	确认当前操作
/od x	/od cancel	取消当前操作
/od em [msg]	—	修改提交信息（/od ps 流程中）
/od ln y	—	接受所有学习提案
/od ln y [N,N]	—	接受指定编号的提案
/od ln x	—	拒绝所有提案
/od ln ad [N] [反馈]	—	调整指定提案

---

B. Core Rules

B.1 Activation & Tool Execution

• OmniDev activates only on /od prefix. Without it, treat as normal conversation.
• First action on any /od message MUST be a tool call — zero text before tools.
• Ad-hoc requests (e.g. /od 这里加个按钮) → use tools to find file, edit code, apply changes directly.
• Image attachments: tool calls FIRST, then explain.

B.2 Workflow Philosophy

• Guided, not forced. Phase order: Blueprint → Plan → Dev → Test → Deploy.
• Phases execute in forward order only, but any phase can be skipped.
• Complexity assessment (S/M/L/XL) provides recommendations, not mandates.

B.3 State File Isolation

• Global: docs/omnidev-state/ (00-project-context.md, metrics.json, config.json)
• Branch-specific: docs/omnidev-state/[branch-name]/ (01-blueprint.md, 02-plan.md, 03-progress.md, 04-design.md, 05-test-report.md, 06-release-notes.md)

B.4 Numbered Quick-Select (数字快捷选择)

At every checkpoint where the user needs to choose an action, present options as a numbered list with Chinese descriptions. The user can reply with:

• The number (e.g. 1, 2, 3)
• The short alias (e.g. /od n)
• The full command (e.g. /od next)

All three forms are equivalent. The AI must parse number replies and map them to the corresponding action.

Format template for all checkpoints:

请选择下一步操作：
  1. 继续下一阶段 (`/od n`)
  2. 修订当前输出 (`/od ad [内容]`)
  3. 跳过某阶段 (`/od sk [阶段]`)
  4. 返回某阶段 (`/od bk [阶段]`)
  5. 执行所有剩余阶段 (`/od al`)

Rules:

• Numbers are context-dependent — each checkpoint defines its own numbered menu (Phase Exit has 5 options, Push has 3 options, etc.).
• If the user replies with a number that is out of range, ask them to choose again.
• Always show the short alias in parentheses next to each option so the user learns the shortcuts over time.
• When interactive_mode is true (see §B.6): Replace the numbered text list with the AskQuestion tool (structured choice UI). The user clicks an option instead of typing a number or command. This saves a request round-trip and reduces token usage.

B.5 Lazy Context Loading

Principle: Do NOT read all state files or scan the entire project upfront. Each phase and command declares exactly what context it needs (see §E Phase Context Requirements). On entering a phase:

1. Read only the files listed in that phase's context_requires block.
2. Skip files that don't exist — their absence is informational (e.g. no 00-project-context.md means onboard hasn't run; trigger §C.1 stack detection instead).
3. Cache across phases within the same session — if a file was already read in a previous phase of the current session, reuse it unless the phase explicitly says reload: true.
4. Never pre-read downstream phase artifacts (e.g. don't read 05-test-report.md during Phase 2).
5. Project scanning is gated: full codebase scans (Grep, Glob, SemanticSearch) are only permitted when the phase's context block includes scan: [target]. Otherwise, operate on the files already known from state files or prior phases.

This keeps token usage proportional to the current phase's actual needs, not the total project size.

B.6 Configuration (config.json)

OmniDev stores user preferences in docs/omnidev-state/config.json. If the file does not exist, treat all settings as their defaults.

{
  "interactive_mode": true,
  "ask_mode_after_od": true,
  "update_source_url": "https://github.com/zayeagle/omnidev-kit.git"
}

Setting	Type	Default	Description
interactive_mode	boolean	true	When true, use the AskQuestion tool to present structured choice UIs at decision points instead of numbered text prompts. Saves requests and tokens.
ask_mode_after_od	boolean	true	When true, enter a Q&A loop after every /od command — present actionable options and accept free-form input so the user stays in an interactive workflow with full tool access until /od x.
update_source_url	string	"https://github.com/zayeagle/omnidev-kit.git"	Remote Git repository URL used by /od up to fetch the latest version. Written automatically during /od install.

Config Commands

• /od cfg — Read and display current config.json (create with defaults if missing).
• /od cfg -i on — Set interactive_mode to true and ask_mode_after_od to true.
• /od cfg -i off — Set interactive_mode to false and ask_mode_after_od to false.

B.7 Interactive Mode

When interactive_mode is true (default) in config.json, the AI MUST use the AskQuestion tool (structured multiple-choice UI) at the following decision points instead of numbered text prompts:

Decision Point	Questions Presented
Phase 0: Complexity Assessment	Confirm/adjust complexity rating (S/M/L/XL); select which phases to include/skip
Phase Checkpoint (after each phase)	Choose next action: continue / adjust / skip / go back
Change Management (/od ch)	Confirm impact assessment: proceed / revise / cancel
Push (/od ps)	Confirm commit message: commit / edit message / cancel
Learning Proposals (/od ln)	For each proposal: adopt / reject / adjust

When interactive_mode is false, use numbered text prompts as defined in §B.4. Do not use AskQuestion.

On first /od activation in a session: Read docs/omnidev-state/config.json (if it exists) to load interactive_mode and ask_mode_after_od. If the file does not exist, assume both are true (default).

B.8 Auto Q&A Loop (自动问答模式)

When ask_mode_after_od is true (default), the AI enters a Q&A loop after every /od command completes. Instead of silently stopping, it presents actionable options AND accepts free-form input, keeping the user in an interactive workflow with full tool access.

Trigger

After every /od command finishes its primary work (phase execution, help display, config change, push, review, learn, update, etc.), the AI MUST present the Q&A prompt as the final action of the response.

Q&A Prompt Format

Use the AskQuestion tool with allow_multiple: false. The question prompt should be "✅ 任务已完成，请选择下一步操作：".

Options are context-adaptive — include only the ones relevant to the current state. Always include other and exit.

id	label	When to show
next_phase	继续下一阶段 (/od n)	A workflow is active and has remaining phases
adjust	修订当前输出 (/od ad)	A phase just completed and its output can be revised
review	代码审查 (/od rv)	Code has been written or modified
test	运行测试 (/od qa)	Code has been written or modified
push	提交并推送 (/od ps)	There are uncommitted changes
learn	自学习与复盘 (/od ln)	A task of M/L/XL complexity just finished
report	生成周报 (/od rp)	Always (low priority, show near bottom)
other	其他（直接输入指令或提问）	Always
exit	结束本次任务 (/od x)	Always

When interactive_mode is false, display the same applicable options as a numbered text list instead of AskQuestion.

Loop Behavior

• User selects a preset option → AI executes the corresponding /od command, then presents the Q&A prompt again.
• User selects "其他" or types free-form text (question, instruction, or any non-/od input) → AI treats it as a continuation of the OmniDev session — OmniDev rules remain active, AI uses tools to fulfill the request, then presents the Q&A prompt again.
• User selects "结束" or types /od x → AI outputs a brief closing summary and stops. The Q&A loop ends.
• User types a new /od command → AI executes it normally, then the Q&A loop continues.

Rules

1. The Q&A prompt is always the very last action — after all output, checkpoints, and phase-specific AskQuestion calls.
2. Every response within the loop also ends with the Q&A prompt (the loop persists until explicit exit).
3. When ask_mode_after_od is false, skip the Q&A loop entirely. The AI completes the /od command and stops.
4. /od cfg -i off disables both interactive_mode and ask_mode_after_od. /od cfg -i on re-enables both.

---

C. Phase Execution Protocol (Auto-Context Loading)

MANDATORY: Before executing ANY phase or special command, you MUST autonomously use the Read tool to load its specific instruction file from .cursor/skills/od/phases/ or .cursor/skills/od/engine/. Do NOT wait for the user to tell you to read it.

Target Phase/Command	File to Read Immediately
Phase 0 / /od onboard	.cursor/skills/od/phases/00-assessment.md
Phase 1 / Phase 2	.cursor/skills/od/phases/01-02-planning.md
Phase 3	.cursor/skills/od/phases/03-development.md
Phase 4 / /od qa	.cursor/skills/od/phases/04-testing.md
/od push, /od change, /od report, /od compress	.cursor/skills/od/engine/special-flows.md
/od learn, /od ln	.cursor/skills/od/engine/evolution.md

After reading the instruction file, follow its context_requires block to load project state files.

C.1 Phase Exit — Checkpoint

After each phase completes, output:

✅ **Phase N 完成: [Phase Name]**
   产出物: [what was produced]
   已加载上下文: [list files/scans actually read]
📍 **当前进度**: 已完成: [...] ✅ | 剩余: [...] ⏳
🔜 **下一阶段: Phase N+1 — [Phase Name]**
   将加载: [list context_requires of next phase]
   将执行: [description]

If interactive_mode is false: Display the numbered list above. Accept number (1–5), alias (/od n), or full command (/od next). STOP — WAIT for user reply.

If interactive_mode is true (default): Use AskQuestion tool to present the same options as a structured choice UI. STOP — WAIT for user selection.

---

D. Project Type Awareness

• Legacy: Follow existing conventions 100%. No forced DDD/TDD. "Write code like a sensible veteran employee."
• Greenfield: Full modern conventions — OpenSpec / TDD / DDD, high test coverage, deployment manifests.

E. MCP Integration

During Phase 3 & 4, proactively use available MCP servers:

• Database MCP: Verify table structures, insert mock data, check data flow.
• Browser MCP (Playwright): Start server → visit page → E2E test → screenshot/DOM verification.
• Discovery: Check .cursor/mcp.json before complex tasks. If critical MCP missing, suggest installation.