daemon inputs: after a hang replaced the only context, the next check through the daemon reuses nothing (#133)

未關閉
#175 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
35/100
Issue 類型
缺陷
描述清晰度
基本清楚
活躍度
活躍
技術堆疊
javascript
領域
testing

研究方向

Read the README's reuse contract, R13, and the daemon and warm-read logic in check.mjs around warmReads and globalKey. Compare the fixed-seed e2e unit's reach on #133 with main, then add an oracle law for reuse of the unedited file's runs after a hang replaces the only context. The issue leaves the cause unconfirmed and describes alternative fixes; done means the law captures the regression and the correct check still matches the cold run.

由索引模型根據 Issue 內容生成。

描述

Summary

On #133 (since at least 9bb7c74a, and at 5fc4b9a4), the fixed-seed e2e unit a check through the daemon that works past the bound reuses no unit in its test-edit check through the daemon. On main (97a4e7a8 and 30928ee8), the same check reuses file b's runs. It reaches incremental check reused evidence and incremental check reused some evidence and recomputed some.

This is a loss of reuse, not of correctness: the check still equals the cold run. The campaign does not lose these states, because 34 and 24 other units reach them. But the unit's own reach moved, and no law requires partial reuse after an edit, so the oracle passes.

Trace

  • #133 at 5fc4b9a4: CI run 44436324847, job 135266004465 (fixed seed 2/5). The unit's reach has a check through the daemon worked past its CLI's bound, an edit that changed the cold result after a hang replaced the only context and edit test changed the cold result, and neither reuse state.
  • Main 30928ee8: run 44435834785, job 135264344204. Its reach also has both reuse states.

Likely cause (plausible, not confirmed)

In this history a slow copy is cut, and a new context warms in place of the cut one. A warm records its reads as global inputs (warmReads, check.mjs), which become part of the unit identity (globalKey). If the replacement warm reads a file or variable that the first warm did not, then the next check's globalKey differs from the one the earlier units were stored under, and every unit misses once.

Possible fixes:

  • A warm in place of a cut context records new reads as hook reads, like a read that arrives after its warm ended.
  • Or the global key covers only the reads of a session's first warm.

Either needs a law that the oracle can judge. One option is to require reuse of the unedited file's runs in the check after a hang replaced the only context, which is a state the grammar already reaches.

Authority

The README's reuse contract and R13 (the daemon is the default). M-30 is met (I8 holds), so this is a speed regression, which belongs to the maintainer's cost rules (P12 / R7).

This issue was written by Isaac.

主要語言
JavaScript
星號
5
分支
1
平均合併
11 小時 18 分鐘
30 天內合併 PR
66

環境準備

這個專案沒有提供開發容器、Dockerfile 或貢獻指南,環境需要你自己搭建:先看它的 README,通用步驟見我們的新手貢獻指南。

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

TanStack/oracles 的其他 Issue

查看 TanStack/oracles 的全部 Issue

相似的 Issue

更多 JavaScript Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。