fix(skills): ship performance diagnosis display metadata

[huangruiteng]

The shipped performance diagnosis skill omitted agents/openai.yaml, so the canonical display-name test fails on current main and packaged hosts receive no explicit display name or invocation prompt. Add the existing UI metadata shape and its setuptools data-file entry so source and wheel installs both expose LoopX Performance Diagnosis.

This is a seven-line packaging correction discovered while triaging #5466 CI. It reuses the other workflow skills' metadata and delivery owner; no new installer logic, invocation policy, profiling behavior, or SQLite policy is introduced. The adjacent simplification pass found no additional abstraction needed.

Written by: model_agent — OpenAI Codex (GPT-6 family).

Validation on head 257c554defa85a2a3a93ff4f48f6e032f2f38f75, base 3cecb0d7cbf3d247cfd996267d7fc64627bd6e49:

• Reproduced the missing-metadata failure on the pinned base; existing metadata and workflow install tests pass after the fix: 27 passed.
• Skill validator, public/private diff inspection and git diff --check: passed.
• Built the frontend with npm ci / npm run build:chat, built a real wheel, and installed it into an isolated environment. The actual installed CLI resolved python_distribution, installed the exact metadata bytes, repeated installation successfully, and removed the managed skill on uninstall. Synthetic usage collection was disabled.
• Exact-scope quality receipt cqr_2e6b98b12ac581b6329c: valid. Native premerge passed its three direct checks; it selected no catalog/risk checks for these two declarative paths, so the focused tests and installed-wheel lifecycle above provide the behavior coverage. No final validation failures/skips or manual holds.

UI scope: host skill-list metadata only; no Dashboard rendering, frontend setting, Lark/API behavior or first-screen layout changes. No local production installation or provider migration is claimed. Other CI failures on #5466 remain separately tracked and are not waived by this fix.

[huangruiteng]

Reviewer: model_agent — GPT-6 family / OpenAI (self-reported)

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

未发现阻塞问题。评审 head：257c554defa85a2a3a93ff4f48f6e032f2f38f75，不可变基线：3cecb0d7cbf3d247cfd996267d7fc64627bd6e49。这是源码与安装包元数据修复；尚未声称用户桌面的技能列表已经更新。

动机

已有 test_packaged_loopx_skills_use_canonical_brand_display_names 要求每个发布的 LoopX 技能明确声明品牌显示名；同一测试在基线因 performance-diagnosis 缺少文件而失败。只加源码文件不能保证 wheel 用户收到它，因此两处声明构成完整修复。能力边界依据 loopx/capabilities/performance_diagnosis/README.md，spec_revision 为 3cecb0d7cbf3d247cfd996267d7fc64627bd6e49；按原文核对 Placement and ownership（implemented：复用管理技能与安装 owner）、Run and read back（implemented：诊断仍须显式调用，不自动安装或执行 profiler）、Complete the diagnosis（implemented：独立基线与本地私有原始证据的要求保留）。显示名的独立判据来自既有测试，未用新增 YAML 反推需求。

改动思路

沿用相邻技能的 interface 元数据和 setuptools data-files，不为七行修复新建安装器或运行状态。Host 发现技能与用户启动诊断是两个步骤：新增名称、短说明、初始提示增加可发现性，但原 SKILL.md、作用域标记及 TS recipe/inspection owner 不变。安装后的文件由现有完整树摘要纳入读回和漂移检查；重复安装与卸载继续由同一 owner 决定。长期维护成本减少了源码与发布包不一致，用户不再收到没有明确名称的技能；无需新增配置、确认或 provider 选择。

具体改动

关键代码讲解

• skills/loopx-performance-diagnosis/agents/openai.yaml:2–4：interface.display_name 为 LoopX Performance Diagnosis，short_description 描述受控本地证据；default_prompt 显式调用该技能，要求 owned target、独立基线与原始 profiles 留在本地，没有授予生产 attach、上传或执行权限。
• pyproject.toml:84–86：将该 YAML 放入 share/loopx/skills/loopx-performance-diagnosis/agents；这是实际 wheel 路径，避免源码测试通过但发行包遗漏。
• 未改的 resolve_workflow_skill_source 与 _install_one_skill：安装态选择 python_distribution，将完整技能树复制到目标 Host 路径，已有摘要、原子替换和 unchanged 分支继续生效。我分别构建并隔离安装了基线和 head wheel，通过真正的 workflow-skills CLI 读回，不借用源码导入或 mock。

对主干的风险

最强反例是 YAML 只在 checkout 存在，wheel 或安装器没交付它。真实包对照已经覆盖：基线安装后缺文件；head 安装后字节与提交文件相同，显示名准确。两臂 SKILL.md、scope、版本标记摘要相同，预览均不创建技能目录，重复安装均为 unchanged；修改已安装技能后，卸载均明确返回非成功并保留修改，恢复标准内容再卸载均删除受管理技能。这也验证未启用诊断时没有新增执行、配额、状态或外部效果。

本地同一组 metadata/workflow 测试：基线 26 passed / 1 expected missing-metadata failure；head 27 passed。技能格式校验、diff 检查通过。首次构建因工作树内旧前端产物被资产门禁拒绝，重建后 head 与基线 wheel 均正常构建；基线重新验证了相同前端源码可使用该构建，未跳过门禁。负例 harness 最初错误期待 modified-uninstall 成功，按原有“不成功且保留修改”的独立合同修正预期后，两臂均通过。没有把这些准备失败当成产品缺陷。没有查询 CI，也未验证真实桌面列表渲染、冻结应用或 profiling 本身；这些不是本次未改边界的新增证据要求。

我的整体评价

APPROVE。当前切片完成了可重放的源文件→wheel→实际安装→读回→安全卸载闭环，long_horizon 与 user_experience 均 improved。新增的是展示信息，复用既有能力名称与权限合同，没有新共享词汇、分类规则或硬义务；语义 advisory 的零候选仅作为补充，完整 diff 与调用方检查才支持此判断。相邻边界的简化审视 found unnecessary：最小正确方案就是现有 owner 的 YAML 和打包项，无须再抽象。后续真实用户安装时应更新管理技能后观察列表；此 APPROVE 不授予合并或本机安装权。

English verdict: APPROVE — 257c554. The seven-line metadata/package fix passes 27 focused tests and real baseline/head wheel install, readback, repeat and safe-uninstall controls; profiling and authority behavior remain unchanged.