test(skills): preserve native Pi loading and session lifecycle

[sowen1023]

Problem

Refs #282, #316, #317.

Scope update, 2026-08-31: following the accepted native-Pi decision, this PR no longer adds inline $skill invocation or hidden Skill-body recovery. The old candidate correctly implemented the earlier requirement; the maintainer has now narrowed that requirement.

Value

Keep Pi as the owner of Skill discovery, invocation, normal history and compaction. Avoid a separate OpenPI body cache, provider-only projection and recovery lifecycle. Users explicitly invoke /skill:name; models can read discovered Skills using the ordinary read tool.

Approach

• Retire the unmerged inline Skill extension, its autocomplete/parser, immutable snapshots, context overlays, compaction reanchoring and dedicated old-contract tests.
• Reuse native Pi as-is: no alias, new loader, recovery marker, setup switch or runtime module.
• Record the scope in Decision 0002 and explain usage and limits in README.
• Add real Pi Session regressions that load the whole OpenPI package with isolated fixture Skills and a deterministic faux provider.
• Relative to current main ba550bc1d918147c326840276eff018b505ed9d8, production extensions, package manifest and lockfile are unchanged. The net PR is documentation and tests. The removal is relative to the previous unmerged candidate, not a removal from shipped main.

Validation

• Node 24.18.0, Bun 1.3.14, Pi 0.84.1.
• bun run check: passed (configuration/discipline checks, format, lint, typecheck).
• Native Skill suite: 10 passed, covering metadata/read, explicit-only Skills, slash expansion and persisted history, direct and streaming steering/follow-up, unknown/missing files, native compaction, no automatic body recovery, and explicit rereading of changed instructions.
• Red/green: with the old extension still loaded, the new ordinary-$skill test fails because hidden bodies are injected; after removal, the complete new suite passes.
• Full bun run test: 1086 Node tests passed, 1 platform-specific skip; 30 Vitest tests passed on the final rerun. The first pre-submit full run hit two existing Subagent watchdog timing failures; both passed in isolation and the full rerun passed without source changes. Those earlier failures are not being hidden or treated as Skill regressions..
• git diff --check and Decision/README local-link checks: passed.
• Hosted CI for the new commit is tracked separately below; old-candidate CI is not evidence for this revision.
• This is deterministic Session integration coverage, not a live paid-model efficacy benchmark or new PTY UI acceptance. Previous inline-completion smoke/review results are historical and do not apply to the replacement.

Impact

• User-visible: keep native slash invocation/completion. $skill is ordinary text, with no OpenPI automatic-loading or inline multi-reference guarantee.
• Model-visible/history: native slash bodies are normal user messages; native reads are normal tool results. Both use Pi persistence and compaction. There is no separate hidden-body overlay.
• Runtime/config/tools: no production delta from main; no new state, settings, tool permissions or dependencies.
• Accepted limit: compaction does not guarantee verbatim Skill retention or automatic rereading. Read-output and total-context limits remain native Pi's limits.
• Compatibility: this intentionally supersedes the unmerged candidate's unchanged-raw-text/hidden-body contract. No existing Session records or user settings are rewritten.
• Scope: no automatic closure of feat: support inline skill references with $ autocomplete #282/feat: add TUI autocomplete for Inline Skill References #317 or the separate feat(skills): support inline $ Skill references #321 PR. Issue test: preserve native Pi Skill invocation and lifecycle #316 records the native contract, with the superseded inline requirements retained as history.
• Merge: no merge or auto-merge enabled. Branch protection requires the current CI checks and an approval from someone other than the latest pusher.

Contributions and history

The earlier work by sowen1023 and bkyssn remains credited in the commit history and linked PRs. Its runtime mechanism is retired because of the maintainer's scope decision, not because those contributors failed the original requirement.

Current revision

• Head: 89f6c88fab44e489be38d276d76c7f9cd63c655f.
• Fast-forward update; contributor commits and history are preserved.
• Required CI and non-last-pusher approval remain separate gates. See the current checks and readiness receipt; this description does not imply approval or merge.

[tt-a1i]

感谢 @bkyssn 和 @sowen1023 的投入。这里由维护者统一整合：以 #322 的 Session 实现为基础，纳入 #321 的 TUI 补全和相关测试，并修复评审发现的队列、压缩与解析边界问题。整合、冲突处理和验证由我们负责，采用的代码会保留原作者贡献与共同署名。请先不用继续重复修改这两块；我们会在整合结果通过验证后合并主 PR，再将另一 PR 标明已整合并关闭。当前尚未合并。

[tt-a1i]

整合提交已推送：e4731fcc90af172b0dbed3325220d42f13cfb81c。在原 #322 上追加提交，没有重写原作者历史。

• 纳入 feat(skills): support inline $ Skill references #321 的补全实现和 Editor 测试；提交保留 Co-authored-by: lwc <1803086257@qq.com>，原 test(skills): preserve native Pi loading and session lifecycle #322 的作者提交保留。
• 修复自动压缩重试丢失 Skill 正文、Unicode 未知名称误匹配，以及补全包装器触发字符组合。
• 保留原始消息、运行期快照、排队输入、原生 slash/file 补全；没有新增工具、权限、配置或 Skill 正文持久化。
• Node 24.18.0 / Pi 0.84.1：bun run check 通过；专项 25/25；bun run test 为 Node 1101 通过、1 个既有平台跳过，Vitest 30/30。
• 两个独立复审视角已通过；第一轮发现的补全组合问题已修复并复审。
• 隔离环境 pi list 仅一个 OpenPI 源，指向本次 checkout；真实 PTY TUI 验证 $inline-smoke、/skill:inline-smoke、@README.md 的 Tab 插入不提交，/reload 后补全仍正常。Provider 行为由真实 Session + 本地 faux provider 验证，未调用外部付费模型。

该提交的 GitHub CI 已全部通过：run 33367106344，包含 Node 22.19.0、Node 24 和 Windows。当前仅剩独立人工审批门禁，尚未合并；维护者正在确认收口方式。合并后将 #321 标明已整合关闭，并在最终 squash 提交保留两位贡献者署名。

[tt-a1i]

原生方案提交与验证回执（2026-08-31）

当前 head：89f6c88fab44e489be38d276d76c7f9cd63c655f；base：ba550bc1d918147c326840276eff018b505ed9d8。

• 已正常快进推送，没有强推或重写贡献者提交历史。
• 净差异只有 README、Decision 0002、Decision 索引和原生 Skill 回归测试；生产 extensions、package manifest、lockfile 与当前 main 一致。
• test: preserve native Pi Skill invocation and lifecycle #316 已同步原生验收边界，旧行内/隐藏正文要求完整保留为历史。不自动关闭 feat: support inline skill references with $ autocomplete #282、feat: add TUI autocomplete for Inline Skill References #317 或 feat(skills): support inline $ Skill references #321。
• 本地 bun run check 通过；全量重跑 1086 项 Node 测试、30 项 Vitest 通过，1 项平台跳过。首轮两项既有 Subagent watchdog 时序失败、单独通过及随后全量通过的情况已在 PR Validation 中保留，没有修改无关代码。
• 本 head 的 CI 全部通过：Node 22.19.0、Node 24、Background terminals (Windows)。两项 Node job 也通过了原生包发现与打包扩展 smoke。标签检查通过。
• 当前无代码冲突；GitHub 仍是 REVIEW_REQUIRED / BLOCKED，不是已经满足全部合并门槛。

剩余门槛：需要另一位有 write 权限、且不是最后推送者的成员 approve 当前版本。 本次没有自批、修改分支保护、开启 auto-merge 或执行合并。模型效果与新终端 UI 验收未新增声明。