bb-spec-review

对规格文档(PRD + arch_spec + MAS)做对抗性评审:完整性 / 一致性 / 可实现性 / 验证覆盖。默认 role=ruthless(规格阶段错误成本最高)。在 MAS frozen 前调用。触发场景:(1) bba-architect 写完 MAS;(2) 显式 /bb-spec-review。

SKILL.md

GitHub

bb-spec-review

职责

跨文档检查 PRD ↔ arch_spec ↔ MAS 一致性、可实现性、验证覆盖;输出严重程度分级 issues。

  • 调用者:bba-architect
  • 关联:bb-challenge-code(通用)
  • 禁止使用:Task / Agent / Skill

Input Args

arg type required 默认 说明
prd_path path true — PRD.md
arch_spec_dir path true — arch_spec 目录
mas_dir path true — spec_mas/(directory containing MAS markdown files)
role enum false ruthless ruthless|linus|balanced
focus string false feasibility,consistency,coverage —
design_name string true — —
stamp string false <auto> —

Output Contract

field 值
artifact_path designs/<name>/spec_review_<stamp>.md
issues_found int
severity_summary {critical,high,medium,low}
consistency_issues list[str]
feasibility_issues list[str]
coverage_issues list[str]
pass bool(critical==0 && high≤2)
valid bool

4-Phase 执行

Phase 1 — build_prompt

prd = open(prd_path).read()
arch = {p.name: p.read_text() for p in Path(arch_spec_dir).glob("*.md")}
mas = {p.name: p.read_text() for p in Path(mas_dir).glob("*.md")}
prompt = render_spec_review_prompt(role, focus, prd, arch, mas)

Phase 2 — run_review

claude --print "<prompt>" > <artifact_path> 或 agent 内部生成。15 min timeout(文档可能很长)。

Phase 3 — parse_spec_review

scripts/parse_spec_review.py:

  • 提 ## Consistency / ## Feasibility / ## Coverage 段
  • 计 severity
  • pass = (critical == 0 && high <= 2)

Phase 4 — return

返回 JSON。pass=false → bba-architect 修 MAS(≤2 iter),仍失败 → escalate 用户。

维度

维度 检查
consistency PRD FR ↔ arch module ↔ MAS FSM 三层一致
feasibility 频率/面积/PDK 能力匹配
coverage verif_plan_seed 完整、corner 覆盖
ambiguity 模糊描述、边界缺失

通过标准

级别 通过
critical == 0
high ≤ 2

收敛 / 失败

状态 行动
pass=true frozen MAS + 开 ready-for-rtl
pass=false & iter<2 bba-architect 修 MAS 重评
iter≥2 escalate 用户

资源索引

  • scripts/build_prompt.py、scripts/run_review.py、scripts/parse_spec_review.py
  • references/spec_review_dimensions.md