chore: Pi 开发依赖跟随最新版并去掉测试里的固定版本号

- pi-hermes-memory / pi-notify / pi-permission-system / pi-subagents 的
  @earendil-works/* devDependencies 改为 latest,pi-rtk-optimizer 放开
  peer 范围;各子包锁文件刷新到 Pi 1.3.0,测试改为验证当前宿主 API。
- 新增 scripts/update-pi-devdeps.sh,统一刷新根与子包锁文件中的 Pi 版本。
- pi-subagents 适配 Pi 1.3:mention clone 改走 systemPromptOverride、
  BOM frontmatter、e2e 读取规范化后的 system 消息与工具列表。
- 根测试不再写死 ssh2 / jiti / better-sqlite3 / context-mode 版本号,
  改为精确版本格式与锁文件一致性断言;新增 latest 约束测试。

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
This commit is contained in:
云服务部-叶林立
2026-10-20 19:38:54 -07:00
co-authored by Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
parent ba164896d2
commit b817d66a11
23 changed files with 4994 additions and 9784 deletions
@@ -112,20 +112,19 @@ describe("subagent usage reaches the parent session's stats (real pi)", () => {
}
});
it("leaves the context-window percentage alone", async () => {
// pi derives context usage from assistant messages only. If that ever
// changed, a delegating session would look like it was filling its context
// with work that happened somewhere else entirely — and users would compact
// for no reason.
it("counts only the result's own text toward context, not its reported usage", async () => {
// Since Pi 1.3 the percentage reflects the transcript's real contents, so
// the message's own few tokens do register. What must not register is the
// reported usage: 150k delegated tokens on a 200k window would read as
// ~75% occupied, and a delegating session would look like it was filling
// its context with work that happened somewhere else entirely.
const session = await realSession();
try {
const before = session.getSessionStats().contextUsage?.percent ?? null;
const pool = new PendingUsagePool();
pool.add({ input: 150_000, output: 400, cacheWrite: 100, cost: 1.5 });
session.sessionManager.appendMessage(toolResultCarrying(pool.drain()) as any);
expect(session.getSessionStats().contextUsage?.percent ?? null).toBe(before);
expect(session.getSessionStats().contextUsage?.percent ?? 0).toBeLessThan(1);
} finally {
session.dispose?.();
}