A reusable Claude Code SKILL for end-to-end software delivery with Claude Code + Codex collaboration.

This SKILL encapsulates a complete software delivery workflow:

需求分析 → 方案 Review → 并行开发+测试规划 → 测试执行循环 → 交付上传

It coordinates Claude Code (development, fixes, self-testing) and Codex CLI (independent review, test execution, issue recording) in a structured delivery pipeline.

• 5-phase delivery pipeline with clear entry/exit criteria
• Dual-agent collaboration: Claude codes, Codex tests
• Defect pool management via TEST-DEFECTS.md (project-local, iteration-isolated)
• Test plan alignment loop: Claude + Codex review cycle (max 20 rounds)
• P0 self-test gate: Claude must pass P0 before handing off to Codex
• Test automation: Convert manual test cases to executable code, reducing token consumption
• Quota resilience: Built-in watchdog for token limit recovery

Copy to your global Claude Code memory directory:

# Create global skills directory
mkdir -p ~/.claude/skills

# Copy the skill file
curl -o ~/.claude/skills/delivery-workflow.md \
  https://raw.githubusercontent.com/maxzsc/delivery-workflow-skill/main/delivery-workflow.md

Then reference it in your global CLAUDE.md or project-level CLAUDE.md:

## Skills
- See ~/.claude/skills/delivery-workflow.md for the full delivery workflow

Copy delivery-workflow.md into your project's skills/ directory.

Activate the SKILL by saying:

按交付流程开始做这个需求
走完整的交付流程
/delivery-workflow

You can also start from any phase:

规划方案已经确认了，直接进入开发和测试阶段
代码写好了，按流程走测试和缺陷管理

1. Codex tests only — Codex executes tests and records issues; Claude does all fixes
2. P0 self-test required — Claude must pass all P0 cases before handing off to Codex
3. Test plan alignment — Claude and Codex iterate until both approve (max 20 rounds)
4. Project-local defect pool — TEST-DEFECTS.md lives in each project, isolated by iteration
5. Test automation — Convert manual tests to code to minimize future token consumption
6. Parallel execution — Coding and test planning run in parallel; fixes and testing run in parallel

MIT