菜单
Run an entire development phase hands-off. Uses a bash wrapper script that invokes /epic_v2 per slice with fresh context. Usage - /phase_v2 <phase-number> [starting-slice]
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/rakheen-dama/b2b-strawman --skill phase-v2命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Implement a specific epic end-to-end — worktree, code, test, PR, review, merge. Supports auto-merge for automated phase execution. Pass the epic number as an argument (e.g. /epic_v2 7).
Generate detailed acceptance test specs for a Kazi feature area. Takes a feature ID from qa/acceptance/INDEX.md, reads phase requirements + architecture, and produces a structured acceptance criteria document with test cases ready for Playwright automation. Usage - /acceptance-spec <feature-id> [--update]
Implement a specific epic end-to-end — worktree, code, test, PR, review, merge. Pass the epic number as an argument (e.g. /epic 7).
正在显示 SKILL.md
基于 SOC 职业分类
| name | phase_v2 |
| description | Run an entire development phase hands-off. Uses a bash wrapper script that invokes /epic_v2 per slice with fresh context. Usage - /phase_v2 <phase-number> [starting-slice] |
Run all remaining slices in a development phase completely hands-off. Each slice gets its own fresh claude invocation via an external bash loop — keeping each context well under the 1M ceiling, even for L-effort slices with large test suites and CodeRabbit iteration.
The orchestration is done by scripts/run-phase.sh, a bash script that:
tasks/phase{N}-*.md)| **NNNX** |)**Done**claude -p "/epic_v2 {SLICE} auto-merge" --dangerously-skip-permissions --model opus
/epic_v2 skill natively supports slice IDs (digits+letter → single-slice mode).Each claude invocation gets a fresh context window. No cross-slice context accumulation; no epic-level prompt-overflow failures.
# Run all remaining slices in phase 10
./scripts/run-phase.sh 10
# Start from a specific slice (skips earlier ones even if not Done)
./scripts/run-phase.sh 10 84A
# Start from the first slice of an epic (e.g., 84 → starts at 84A)
./scripts/run-phase.sh 10 84
# Preview which slices will run (no execution)
./scripts/run-phase.sh 10 --dry-run
# Watch progress in real-time
tail -f tasks/.phase-10-progress.log
# Check which slices are done
grep 'Done' tasks/phase10-invoicing-billing.md | head -20
If a slice fails (review found unfixable issues, build errors, etc.):
tail -50 tasks/.phase-10-progress.log/epic_v2 {SLICE} interactively./scripts/run-phase.sh 10
(it will skip all Done slices automatically)| Aspect | /phase (v1) | /phase_v2 |
|---|---|---|
| Orchestrator | Claude agent (single session) | Bash script (external loop) |
| Unit of work per claude invocation | Whole phase | Single slice |
| Context limits | Hits ~75%, needs manual restart | Fresh 1M context per slice |
| Merge approval | Manual (asks user) | Auto-merge after review passes |
| Failure handling | Agent tries to recover in-session | Script stops, user restarts |
| Monitoring | Watch Claude output | tail -f on log file |
| Parallel slices | No | No (sequential is safer) |
Earlier versions of this script grouped slices by epic and passed /epic_v2 {EPIC}, letting one invocation handle all slices in the epic as a single PR. That broke down for large epics (L-effort slices + 4674-test backend suite + CodeRabbit iteration) — a single /epic_v2 invocation could exhaust its 1M context window before finishing review cycles. Slice-level invocation keeps each context comfortably bounded.
Trade-off: more PRs (one per slice instead of one per epic). If multi-slice bundling is desirable for a specific phase (e.g., tightly coupled DDL + code), invoke /epic_v2 {EPIC} manually outside the script.