ai-dev-loop（AI 开发循环）

AI Development Loop

Autonomous task execution with QA gates across multiple AI platforms.

Overview

The AI Development Loop enables fully autonomous feature development where:

• AI agents pick up and implement tasks from a queue
• You do QA only (approve or reject in Testing column)
• Multiple platforms (Claude CLI, Cursor, Codex) can work in parallel
• Rate limits are maximized by switching between platforms

Architecture

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│   BACKLOG   │────▶│   TO DO     │────▶│  TESTING    │────▶│    DONE     │
│             │     │             │     │             │     │             │
│  PRDs ready │     │ Agent picks │     │ YOU review  │     │  Shipped    │
│             │     │ & builds    │     │ & approve   │     │             │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘
                          │                   │
                    ┌─────┴─────┐       ┌─────┴─────┐
                    │  Claude   │       │  Reject   │
                    │  Cursor   │       │  → To Do  │
                    │  Codex    │       └───────────┘
                    └───────────┘

Task Lifecycle

1. Task Creation

Tasks live in .agents/TASKS/[task-name].md with structured metadata:

## Task: [Feature Name]

**ID:** feature-name-slug
**Status:** Backlog | To Do | Testing | Done
**Priority:** High | Medium | Low
**PRD:** [Link](../PRDS/prd-file.md)

### Agent Metadata

**Claimed-By:** [platform-session-id]
**Claimed-At:** [timestamp]
**Completed-At:** [timestamp]

### Progress

**Agent-Notes:** [real-time updates]
**QA-Checklist:**

- [ ] Code compiles/lints
- [ ] Tests pass (CI)
- [ ] User acceptance
- [ ] Visual review

### Rejection History

**Rejection-Count:** 0
**Rejections:** [list of rejection notes]

2. Task Claiming

When an agent runs /loop:

1. Scans .agents/TASKS/ for Status: To Do
2. Sorts by priority (High > Medium > Low)
3. Skips tasks with active claims (< 30 min old)
4. Updates task with Claimed-By and Claimed-At

3. Implementation

Agent works on the task:

1. Reads task file and linked PRD
2. Checks .agents/SESSIONS/ for related past work
3. Implements the feature/fix
4. Updates Agent-Notes with progress
5. Creates branch and commits

4. Quality Check

Before moving to Testing:

1. Runs qa-reviewer skill
2. Updates QA-Checklist items
3. Ensures code compiles/lints

5. Completion

Agent finalizes:

1. Sets Status: Testing
2. Sets Completed-At timestamp
3. Adds final summary to Agent-Notes
4. Prompts for next action

6. QA Gate (Your Turn)

In Kaiban.md:

1. Review Testing column
2. Click task to see PRD preview
3. Check linked PR
4. Approve: Drag to Done
5. Reject: Click reject, add note → returns to To Do

7. Rejection Handling

When rejected:

1. Status returns to To Do
2. Rejection-Count increments
3. Rejection note added to history
4. Next /loop picks up with full context

Multi-Platform Strategy

Platform Strengths

Parallel Execution

Multiple platforms can work simultaneously:

• Each claims different tasks
• Claims prevent conflicts (30-min lock)
• Shared state via task files

Rate Limit Handling

When rate limited:

1. Agent saves progress to Agent-Notes
2. Releases claim (clears Claimed-By)
3. Suggests switching platform
4. User continues with different platform

Daily Workflow

Morning QA Session

1. Open Kaiban.md extension in VS Code
2. Review Testing column
3. Approve good work → Done
4. Reject with notes → To Do

Throughout Day

# Claude CLI
claude
> /loop   # Process task
> /loop   # Next task
# Rate limited? Switch to Cursor

Rate Limit Strategy

Claude limit? → Switch to Cursor
Cursor limit? → Switch to Codex
All limited? → QA time (review Testing)

Integration Points

Kaiban.md Extension

• Visual Kanban board for .agents/TASKS/
• Drag & drop status changes
• PRD preview panel
• Reject button with note input
• Agent claim status badges

Existing Skills

• qa-reviewer: 6-phase quality verification
• session-documenter: Auto-document completed work
• rules-capture: Learn from rejection feedback

Git Workflow

• Branch per task: feature/[task-id]
• Commits with clear messages
• PR linked in task file

Not a Daemon

Important: /loop is NOT a background process.

• Each invocation handles ONE task
• Returns control to user
• User decides to continue or stop
• Respects "never run background processes" rule

Claim Expiration

Claims expire after 30 minutes:

• Handles agent crashes
• Handles rate limit interruptions
• Previous Agent-Notes provide context for pickup
• Enables multi-platform handoff

Best Practices

For Task Creation

• Write clear, actionable task descriptions
• Link to PRD for requirements
• Set appropriate priority
• Include testing criteria

For Agents

• Read task and PRD thoroughly before starting
• Update Agent-Notes regularly
• Run qa-reviewer before completing
• Create clean, focused commits

For QA (You)

• Review PRD alongside implementation
• Provide specific rejection feedback
• Approve incrementally (don't batch)
• Keep Testing column short