SKILL.md

stars:0
forks:0
watches:0
last updated:N/A

dev-flow - One-stop Code Change Workflow


Core Design: Check Agent to Avoid Self-Bias

Use independent Check Agent for code review phase. Main Agent handles coding and coordination.

AgentResponsibilityLifecycle
Check AgentCode review, checking modified files + affected filesStart BEFORE coding → Release when outputting results

Mandatory Requirements

RequirementDescription
MUST NOT pause midwayAfter user confirmation, MUST execute all steps continuously
MUST NOT skip testingCompilation/testing phase MUST be executed
Check Agent MUST start BEFORE codingStart before coding, NOT after
ROLLBACK MUST reuse AgentAfter rollback, MUST use SendMessage to reuse Check Agent, MUST NOT create new one

Phase 1: Requirement Confirmation

Flow

Main Agent reads code → Analyzes gaps using checklist → AskUserQuestion → User answers → Continue or end

Main Agent MUST:

  1. Read related code to understand current implementation
  2. Check against gaps checklist in requirement.md
  3. Call AskUserQuestion with N question objects
  4. After receiving answers, if gaps remain, continue asking
  5. If no gaps, output requirement summary and ask user to confirm

[MANDATORY] AskUserQuestion format: see reference/requirement.md


Phase 2: Coding

Flow

Start Check Agent (BEFORE coding!)
    ↓
Main Agent performs coding

[MANDATORY] Check Agent MUST start BEFORE coding begins.


Phase 3: Check

Flow

SendMessage(Check Agent) → Check code issues
    ↓
┌───────────────────────────────────────────────┐
│ Check Result?                                 │
│   ├─ No issues → Enter compilation testing    │
│   └─ Has issues → ⏪ ROLLBACK to coding       │
└───────────────────────────────────────────────┘

[MANDATORY] SendMessage MUST include the list of modified files.

See: reference/validation.md


Phase 4: Compilation Testing

Flow

Main Agent executes compilation testing
    ↓
┌───────────────────────────────────────────────┐
│ Test Result?                                  │
│   ├─ Pass → Enter output results              │
│   └─ Fail → ⏪ ROLLBACK to coding             │
└───────────────────────────────────────────────┘

[MANDATORY] MUST NOT skip testing.

See: reference/validation.md for test commands


Phase 5: Output Results

Release Check Agent
    ↓
Output task complete

Rollback Mechanism

Check/Test fail → ⏪ ROLLBACK to coding → SendMessage(Check Agent_ID) → Re-execute check→test

See: reference/validation.md for rollback declaration format


User Commands

CommandEffect
/abortAbort task
够了/确认/开始/OKStop questioning, start coding

Agent Invocation Rules

RuleDescription
Before calling AgentMUST NOT output text, call directly
After calling AgentWait for result directly, MUST NOT output "waiting" text
If AskUserQuestion returns duplicate answersUse only the first answer, ignore duplicates
SendMessageMUST include list of modified files
Agent releaseRelease after entire workflow completes

See: reference/agent-templates.md


Resource Navigation


$ARGUMENTS

    Good AI Tools