#!/usr/bin/env python3 """Build the local, inspectable Skill-Up management dashboard.""" from __future__ import annotations import html import json import re import subprocess from pathlib import Path ROOT = Path(__file__).resolve().parents[1] OUTPUT = ROOT / "work/2026-08-12-alibaba-skill-up-adoption/skill-up-management-dashboard.html" RUN_ROOT = ROOT / "work/2026-08-12-alibaba-skill-up-adoption/runs" def text(value: object) -> str: return html.escape(str(value), quote=True) def first_scalar(path: Path, key: str) -> str: pattern = re.compile(rf"^{re.escape(key)}:\s*(.+?)\s*$") for line in path.read_text(encoding="utf-8").splitlines(): match = pattern.match(line) if match: return match.group(1).strip().strip('"\'') return "" def relative(path: Path) -> str: return path.relative_to(ROOT).as_posix() def skill_up_version() -> str: try: result = subprocess.run( ["skill-up", "--version"], cwd=ROOT, check=True, capture_output=True, text=True, timeout=10, ) return result.stdout.strip() except (FileNotFoundError, subprocess.SubprocessError): return "skill-up command unavailable" def load_suites() -> list[dict[str, object]]: suites: list[dict[str, object]] = [] for eval_path in sorted(ROOT.glob("standards-stack/agent-skills/skills/*/evals/eval.yaml")): skill_dir = eval_path.parents[1] cases = [] for case_path in sorted((eval_path.parent / "cases").glob("*.yaml")): cases.append( { "id": first_scalar(case_path, "id"), "title": first_scalar(case_path, "title"), "description": first_scalar(case_path, "description"), "path": relative(case_path), } ) suites.append( { "skill": skill_dir.name, "eval_path": relative(eval_path), "cases": cases, } ) return suites def load_runs() -> list[dict[str, object]]: runs: list[dict[str, object]] = [] if not RUN_ROOT.exists(): return runs for result_path in sorted(RUN_ROOT.glob("*/iteration-*/result.json")): payload = json.loads(result_path.read_text(encoding="utf-8")) case_results = payload.get("case_results", []) statuses = [str(item.get("status", "UNKNOWN")) for item in case_results] status = statuses[0] if len(set(statuses)) == 1 and statuses else "MIXED" report_path = result_path.with_name("report.html") runs.append( { "name": result_path.parents[1].name, "iteration": result_path.parent.name, "status": status, "cases": len(case_results), "started": payload.get("start_time", ""), "report": relative(report_path) if report_path.exists() else "", "result": relative(result_path), } ) return sorted(runs, key=lambda item: str(item["started"]), reverse=True) def build() -> str: suites = load_suites() runs = load_runs() case_count = sum(len(suite["cases"]) for suite in suites) status_counts = {name: sum(1 for run in runs if run["status"] == name) for name in ("PASS", "FAIL", "ERROR")} suite_sections = [] for suite in suites: case_rows = "".join( f""" {text(case['id'])} {text(case['title'])}{text(case['description'])} 查看 YAML """ for case in suite["cases"] ) suite_sections.append( f"""
TEST SUITE

{text(suite['skill'])}

{len(suite['cases'])} cases

{text(suite['eval_path'])}

{case_rows}
Case ID目的定义
""" ) run_rows = "".join( f""" {text(run['name'])}{text(run['iteration'])} {text(run['status'])} {text(run['cases'])} {text(str(run['started']).replace('T', ' ')[:19])} {f'HTML' if run['report'] else '—'} · JSON """ for run in runs ) return f""" Skill-Up 测试管理台
AGENT SKILL EVALUATION · ZHCTPROMPT

从真实问题，到失败报告，再到修复回归

这是 Skill-Up 在协作空间里的统一管理入口。先看用例，再看运行报告；失败只允许归因到 Skill、Eval 或 Engine，对应层修复后先跑失败用例，再跑受影响套件。

开始使用打开 Markdown 教程打开 Jack Fork
{len(suites)}测试套件
{case_count}已登记用例
{len(runs)}历史运行记录
{status_counts['PASS']}PASS 运行

我到底在哪管理？

Skill-Up 当前是“仓库即管理台”，没有独立 SaaS 后台。日常只需记住下面三个入口。

① 用例库

skills/<skill>/evals/

维护输入、通过条件、禁止结论和用例清单。

② 运行报告

work/.../runs/

查看每次 PASS、FAIL、ERROR、回答和 HTML 报告。

③ 团队治理

skill-up-evaluation-regression

管理失败分类、修复顺序、安全边界、任务与 review。

测试用例库

用例已通过配置验证；“已登记”不等于“全部跑过模型”。

{''.join(suite_sections)}

运行记录

保留红灯是证据，不覆盖旧目录。PASS/FAIL 表示已得到回答并评分；ERROR 表示引擎基础设施没有完成公平评测。

{run_rows}
运行状态Case开始时间报告

五分钟上手

看版本确认本机命令
列用例list-cases
验配置validate
试运行dry-run
跑单例include-case
看报告report.html

1. 进入项目并看版本

cd /path/to/zhctprompt
skill-up --version

{text(skill_up_version())}

2. 查看并验证套件

skill-up list-cases standards-stack/agent-skills/skills/binzhou-one-card-regression-gate/evals/eval.yaml
skill-up validate standards-stack/agent-skills/skills/binzhou-one-card-regression-gate/evals/eval.yaml
skill-up run standards-stack/agent-skills/skills/binzhou-one-card-regression-gate/evals/eval.yaml --dry-run

3. 只跑一个用例

skill-up run standards-stack/agent-skills/skills/binzhou-one-card-regression-gate/evals/eval.yaml \
  --include-case-name recharge-arrival-not-closed \
  --output-dir work/2026-08-12-alibaba-skill-up-adoption/runs/regression-recharge-YYYYMMDD \
  --format html --iteration 1
关键：每次使用新的 output-dir。失败目录不能覆盖；修复后先重跑失败 case，再跑整个相关套件。

失败怎么管

SKILL_DEFECT

Skill 路由或业务判断错误。修改 Skill，不碰正确断言。

EVAL_DEFECT

回答正确但断言过死，或错误答案能漏过。修改 case judge。

ENGINE_INFRA

网络、Codex、凭据、超时、runtime 出错。修引擎，不误伤业务 Skill。

团队责任

产品/交付

真实问题和验收边界

测试

正负与边界场景

Agent

YAML、运行、初分

开发/维护人

修正确认层

Reviewer

防弱化、防越界

刷新这个管理台

新增用例或运行报告后，在仓库根目录执行：

python3 scripts/build-skill-up-management-dashboard.py

然后验证生成页，再提交协作空间。页面每次从当前 case YAML 与 result.json 重建，不单独维护第二份状态。

安全边界

默认只使用本机 runtime、Codex 单轮隔离和本地规则 judge。OpenSandbox、远程 judge、MCP、OTLP、自定义 HTTP engine、沙箱绕过以及客户/生产敏感数据都不默认启用。Skill-Up PASS 证明 Agent Skill 行为，不自动等于生产支付、硬件或第三方资金链 E2E。
""" def main() -> None: OUTPUT.parent.mkdir(parents=True, exist_ok=True) OUTPUT.write_text(build(), encoding="utf-8") print(f"built {relative(OUTPUT)}") if __name__ == "__main__": main()