test: preserve native Pi Skill invocation and lifecycle

[bkyssn]

当前范围（2026-08-31 修订）

根据 Discussion #324 的已确认决定，本 Issue 改为复用 Pi 原生 Skill 调用与生命周期，不增加 OpenPI 的行内加载和正文恢复层。对应 PR #322；规范记录为 Decision 0002。

这是维护者主动调整需求，不是原贡献者没有完成旧要求。旧需求保留在文末作为历史，不再作为当前验收依据。

当前合同

• 模型自主选择 Skill：Pi 提供目录，正文通过原生 read 按需读取。
• 用户显式调用：使用原生 /skill:name；补全、正文展开及 steering/follow-up 继续由 Pi 负责。
• 正文通过普通用户消息或工具结果进入正常 Session 历史，压缩与恢复完全交给 Pi。
• OpenPI 不提供专门的 $skill 语法、独立正文缓存、provider-only 隐藏投影、压缩后补回正文或保活重读入口。
• 接受原生限制：压缩后不保证全文保留，也不保证模型自动重读；原生输出与上下文容量限制仍然适用。

当前验收

• PR 候选撤掉行内解析、补全适配、正文快照、上下文注入及压缩后重新关联；未新建替代运行时模块。
• README 和 Decision 说明原生用法、持久化差异与限制。
• 测试加载整个 OpenPI 包，证明 $skill 不再触发程序注入，Pi 目录和原生读取仍可用。
• 测试证明原生斜杠调用、explicit-only Skill、直接及 streaming steering/follow-up、正常历史保存可用。
• 测试覆盖未知调用、文件读取失败、真实 Pi 压缩不额外补回正文、后续显式调用读取更新后的文件。
• 本地完整检查与测试通过；不是模型遵循能力或新终端 UI 验收。
• 当前 PR head 89f6c88fab44e489be38d276d76c7f9cd63c655f 的必需 CI 检查通过：Node 22.19.0 / Node 24 / Windows。
• 取得非最后推送者的有效 approve，满足仓库合并门槛。

以上勾选表示候选代码/本地验证，不代表已经合并或发布。CI 状态随 #322 当前提交核对。Parent #282、#317 与 PR #321 没有因此自动关闭。

---

历史需求：行内 $skill＋隐藏正文（已被上述原生方案替代）

Parent

• feat: support inline skill references with $ autocomplete #282

What to build

Make $skill-name a working Inline Skill Reference for every normal Pi Session input path.

When a user submits a known reference, keep the raw user-visible message unchanged and add the referenced Skill content as hidden model-visible context. Resolve references only from the Skills that Pi supplies for that model turn. Pi remains responsible for Skill discovery, trust, collisions, diagnostics, and reload.

Acceptance criteria

• A known $skill-name reference in a submitted message adds the corresponding frontmatter-stripped Skill content to model-visible context.
• The raw user-visible message remains unchanged.
• Expansion works at the shared Session boundary used by TUI, RPC, print mode, and other normal input paths.
• $ begins a reference only at the start of input or after horizontal whitespace.
• Normal punctuation can follow a known Skill name without becoming part of the reference.
• Text such as abc$skill-name remains ordinary text.
• An escaped \$skill-name remains unchanged and is not expanded.
• An unknown $skill-name remains unchanged and adds no Skill content.
• Each distinct referenced Skill is loaded once, in first-reference order.
• References resolve only against Pi's Skills for the current model turn; OpenPI does not scan directories or maintain a Skill registry or cache.
• A Skill read failure is surfaced through Pi's extension error path and is not reported as successfully loaded.
• Tests separately prove raw user-visible text and hidden model-visible Skill content.
• The repository's complete check and test commands pass from the current main branch.

Blocked by

• None — can start immediately.

[sowen1023]

I would like to work on this issue. I will implement the shared Session-level $skill-name expansion at Pi’s before_agent_start seam, resolve only against the Skills supplied by Pi for the current turn, preserve the raw user-visible prompt, inject deduplicated hidden Skill content in first-reference order, surface read failures through the extension error path, and add focused regressions for lexical boundaries, ordering, unknown/escaped references, and projection separation. I will keep Pi as the sole Skill registry and run the full repository check and test suites.

[bkyssn]

感谢认领。这个子 Issue 的实现已经由 #321 覆盖，相关代码、测试和 CI 均已完成。为避免重复开发，请暂缓继续实现；#321 合并后本 Issue 会自动关闭。