feat(skill): V1.21.0-alpha SkillRunner 基础设施 + pentest-coordinator 首 skill LLM-driven

[Wool-xing]

Summary

skill rollout 起点 — V1.x rollout 收尾后下个里程碑。新加 SkillRunner 基础设施 (与 AgentRunner 平行) + pentest-coordinator 首 skill 落地, 解锁 14 rollout skill 后续流水线。

• skill 15/32 active (8 production + 7 script + 15 rollout + 2 vision) — V1.20 14 → V1.21 15
• 基础设施一次落地, 后续每 PR ≈ 1 skill 推进
• 选 pentest-coordinator 作首 skill 理由: 编排型价值清晰 (LLM 读 PRD + 决定 5 子 skill 顺序 + 授权检查), 与 V1.19 pentest-tester expert 配对复用上下文

基础设施变更

• runtime/orchestrator/agents/base.py 加 SKILL_RUNNERS dict + register_skill deco + get_skill_runner (镜像 AGENT_RUNNERS pattern)
• SkillRunner 复用 AgentRunner 抽象 (接口 100% 同), 仅 registry 独立避免 expert/skill 同名冲突
• runtime/orchestrator/skills/__init__.py 新建 (镜像 agents/init.py)
• runtime/orchestrator/skills/pentest_coordinator.py 新建首 skill
• runtime/orchestrator/adapters/experts.py kind=="skill" 分支接 skill runner lookup
  • 路径优先级: catalog hard-block (rollout/vision/unknown) → expert runner → skill runner (新) → SCRIPT_MAP fallback

pentest-coordinator schema

LLM 读 PRD + 上游 pentest-tester expert 产物 → 5 阶段并发编排计划:

1. recon (单线 20min) → pentest-recon skill
2. vuln_assessment (5 攻击域并发 45min) → pentest-vuln skill
3. exploit (5 攻击域并发 120min) → pentest-exploit skill
4. post_exploit (optional 60min)
5. report (单线 15min) → pentest-report skill

含 authorization_check (tagent.yml pentest.authorized + scope 必须显式) + subagent_pool (max=5 + aux_client 隔离) + refuse_conditions (4 条) + Shannon 哲学过滤 (仅 working PoC 入报告)。

变更

• runtime/orchestrator/agents/base.py SKILL_RUNNERS 基础设施
• runtime/orchestrator/skills/__init__.py + pentest_coordinator.py 新建
• runtime/orchestrator/adapters/experts.py kind=skill 接 runner
• 03-技能定义/pentest-coordinator.md SKILL_IMPL_STATUS rollout → production
• runtime/tests/test_skill_runners.py 新建 (照 test_agent_runners.py pattern)
• runtime/tests/test_impl_status_filter.py skill 数字 sync (8/7/15/2) + rollout 测试改 pentest-recon
• README + ROADMAP 数字 + V1.21 Changelog

Test plan

• pytest test_agent_runners.py — 34/34 PASS (expert layer 无 regression)
• pytest test_skill_runners.py — 4/4 PASS (1 skill × 3 维度 + 1 sync 锚)
• pytest test_impl_status_filter.py — 13/13 PASS (X4 防 mock 闭环不退化)
• sanity import: from runtime.orchestrator.skills import get_skill_runner → 'pentest-coordinator' registered, mock_output schema OK, summary "渗透编排 5 阶段 / 并发上限 5 / 拒绝条件 4"
• pre-commit (ruff / detect-secrets / 路径守 / agents-skills-utils 数量校验 / markdownlint) PASS
• CI 13 jobs 待 PR 触发

下一步

继续 skill rollout 流水线 (V1.22+):

• V1.22 candidate: mobile-test 或 system-test (跟对应 expert 配对的细分用例 skill)
• 或: pentest-recon / pentest-vuln / pentest-exploit / pentest-report (复用 pentest-coordinator 上下文)