dev-flow - One-stop Code Change Workflow

Core Design: Sub Agent Processing to Avoid Self-Bias

Use independent sub Agents for critical phases. Main Agent ONLY handles coding and coordination.

Mandatory Requirements

Phase 1: Requirement Confirmation

Flow

Step 1: Start Questioning Agent
    ↓
Step 2: Get user requirements → Summarize requirements
    ↓
Step 3: Loop
    ↓
SendMessage(Questioning Agent) → Find gaps
    ↓
┌───────────────────────────────────────────────┐
│ Questioning Agent Result?                     │
│                                               │
│   Has gaps:                                   │
│     1. Output brief explanation (found N issues)│
│     2. AskUserQuestion                        │
│     3. Receive answer → Summarize → SendMessage → Loop│
│                                               │
│   No gaps:                                    │
│     1. Output "Questioning Agent found no new gaps"│
│     2. AskUserQuestion(supplement?)           │
│     3. Has supplement → Merge summary → Go to Step 2│
│     4. No supplement → End loop               │
│                                               │
│   User halt → End loop                        │
└───────────────────────────────────────────────┘
    ↓
Release Questioning Agent

[MANDATORY]After receiving Agent result, MUST output brief explanation first, then call AskUserQuestion. Order: Output explanation → AskUserQuestion.

[MANDATORY]After receiving AskUserQuestion answers, MUST SendMessage to let Questioning Agent continue finding gaps. MUST NOT skip directly to asking for supplement.

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 (Code + Test Plan Validation)

Check Agent MUST validate two things: (1) code correctness, (2) test plan completeness

Flow

SendMessage(Check Agent) → Check code + Validate test plan
    ↓
┌─────────────────────────────────────────────────────────────┐
│ Check Result?                                               │
│   ├─ Code issues → ⏪ ROLLBACK to coding (fix code)         │
│   ├─ Test plan incomplete → ⏪ ROLLBACK to coding (add tests)│
│   └─ Both passed → Enter browser testing                    │
└─────────────────────────────────────────────────────────────┘

Validation rules:

Check Agent evaluates whether the test plan (from Phase 4) adequately covers all modifications. Use reasonable judgment:

• A simple typo fix may need only one test
• A new function needs success + failure paths
• A UI change needs interaction + edge cases
• A backend API needs validation + error handling

Check Agent output should include:

• Test coverage assessment (sufficient / needs improvement)
• Specific missing test cases (if any)
• Recommendation (proceed / add tests / rethink design)

If test plan is incomplete, Check Agent MUST explain what's missing and why.

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

See: reference/validation.md

Phase 4: Browser Testing

[MANDATORY]Before starting Phase 4, confirm Check Agent has validated the test plan and result is PASS. If Check Agent returned warnings (missing exception paths), document them in test report but may proceed.

Flow

Confirm Check Agent validation passed
    ↓
┌─────────────────────────────────────────────────────┐
│ Check agent-browser skill availability              │
│   ├─ Available → Proceed to browser testing         │
│   └─ Not available → Fallback: skip testing         │
└─────────────────────────────────────────────────────┘
    ↓
┌─────────────────────────────────────────────────────┐
│ Testing Mode?                                        │
│                                                       │
│   Browser testing available:                         │
│     Main Agent identifies test coverage              │
│     ↓                                                │
│     Create test cases → Execute with agent-browser   │
│     ↓                                                │
│     Output test report → Pass/Fail decision          │
│                                                       │
│   Fallback (agent-browser unavailable):              │
│     Output fallback notice to user                   │
│     ↓                                                │
│     Provide manual test instructions                 │
│     ↓                                                │
│     AskUserQuestion: user confirms manual test       │
│     ↓                                                │
│     Enter output results                             │
└─────────────────────────────────────────────────────┘

[MANDATORY]If agent-browser available: MUST NOT skip testing.

[MANDATORY]If agent-browser unavailable: MUST output fallback notice + manual test instructions, then AskUserQuestion for user confirmation.

[MANDATORY]Test coverage: Each modification MUST have corresponding test case.

[MANDATORY]Output test report: MUST show test cases and results to user.

See: reference/validation.md for test commands, browser testing guide, and fallback procedure

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

Agent Invocation Rules

See: reference/agent-templates.md

Resource Navigation

• Requirement Confirmation Rules: reference/requirement.md
• Check→Test Rules: reference/validation.md
• Agent Templates: reference/agent-templates.md

$ARGUMENTS