Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs(exp-wine): record measured warm-cache result and the queued 8-core leg

Browse Source
This commit is contained in:
Tianyi Cui
2026-07-27 12:39:29 +08:00
parent 38eb521e00
commit b052cd1161
3 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
.agents/notes/proposed/process/2026-07-27-wine-windows-gates-experiment.i18n.yaml
View File

@@ -2,5 +2,5 @@
# side as of the last confirmed-consistent state. Both languages carry equal authority; # side as of the last confirmed-consistent state. Both languages carry equal authority;
# after editing either side, bring the other along and re-record with: # after editing either side, bring the other along and re-record with:
# pnpm run verify-translation-pairing --write .agents/notes/proposed/process/2026-07-27-wine-windows-gates-experiment.md # pnpm run verify-translation-pairing --write .agents/notes/proposed/process/2026-07-27-wine-windows-gates-experiment.md
2026-07-27-wine-windows-gates-experiment.md: 9e2db947eceee7e3e2fee63f8fe2ac90de1cd13d 2026-07-27-wine-windows-gates-experiment.md: 47a37ddb48f4321f916c7f7a0cb96ae80b133103
2026-07-27-wine-windows-gates-experiment.zh.md: a4b938faa6ae27bd068db9a952ebb1432ec7ca3f 2026-07-27-wine-windows-gates-experiment.zh.md: 3a912861110a06b39bfb2c37395fc6a061bdfbe6

2
.agents/notes/proposed/process/2026-07-27-wine-windows-gates-experiment.md
View File

@@ -18,6 +18,8 @@ Dependencies install natively on Linux with `supportedArchitectures` extended to
The lane targets the wall clock of the Linux CI jobs (about two minutes), from four levers: the master-refreshed pnpm store cache (restore-only, same key as ci.yml), Wine provisioning (apt install, Windows Node download, `wineboot`) running concurrently with `pnpm install`, the two blocking surfaces running concurrently — the same shape `run-gates` gives them on native Windows — and an apt-archive cache keyed on the runner image so Wine's package downloads are paid once per image version. The lane targets the wall clock of the Linux CI jobs (about two minutes), from four levers: the master-refreshed pnpm store cache (restore-only, same key as ci.yml), Wine provisioning (apt install, Windows Node download, `wineboot`) running concurrently with `pnpm install`, the two blocking surfaces running concurrently — the same shape `run-gates` gives them on native Windows — and an apt-archive cache keyed on the runner image so Wine's package downloads are paid once per image version.
Measured on 2026-07-27: 2m46s end-to-end on a warm-cache pull-request run (setup and cache restores ~17s, concurrent install+provision 33s, concurrent gates 110s), against 1.5–2.5 minutes for the Linux CI jobs and 7–9 minutes for the paid Windows lane; a cold-cache run pays roughly one extra minute. The 8-core benchmark leg never left the queue — the restricted `dsh-ubuntu-*` pools were also observed queueing indefinitely from the sibling KVM experiment — so the standard-runner number stands as the result, and no larger box is needed to hit the target.
This is deliberately a fidelity probe, not a drop-in replacement: Wine reimplements the Win32 API over a case-sensitive ext4 (NTFS case-insensitivity is not emulated by default), provides no ConPTY, and substitutes its own security-descriptor and `MoveFileExW` semantics — exactly the surfaces the repo's `win32.ts` modules and PTY backend care about. The experiment measures which blocking gates pass, which fail for Wine reasons rather than product reasons, and the wall-clock cost relative to the recorded Windows benchmark lanes. This is deliberately a fidelity probe, not a drop-in replacement: Wine reimplements the Win32 API over a case-sensitive ext4 (NTFS case-insensitivity is not emulated by default), provides no ConPTY, and substitutes its own security-descriptor and `MoveFileExW` semantics — exactly the surfaces the repo's `win32.ts` modules and PTY backend care about. The experiment measures which blocking gates pass, which fail for Wine reasons rather than product reasons, and the wall-clock cost relative to the recorded Windows benchmark lanes.
Promotion, if the verdict is positive: fold the Wine lane in as the pull-request Windows signal for blocking gates and demote the real-Windows pool to the master serial reference; otherwise record the failure class here and keep the pool. Promotion, if the verdict is positive: fold the Wine lane in as the pull-request Windows signal for blocking gates and demote the real-Windows pool to the master serial reference; otherwise record the failure class here and keep the pool.

2
.agents/notes/proposed/process/2026-07-27-wine-windows-gates-experiment.zh.md
View File

@@ -18,6 +18,8 @@ Pull request 的 Windows 通道存在的意义是证明两个阻断性 win32 表
该通道以 Linux CI 作业的墙钟（约两分钟）为目标，靠四个杠杆：master 刷新的 pnpm store 缓存（只恢复，与 ci.yml 同键）、Wine 供给（apt 安装、Windows Node 下载、`wineboot`）与 `pnpm install` 并发运行、两个阻断表面并发运行——与 `run-gates` 在原生 Windows 上给它们的形状相同——以及按 runner 镜像为键的 apt 归档缓存，使 Wine 的包下载每个镜像版本只付一次。 该通道以 Linux CI 作业的墙钟（约两分钟）为目标，靠四个杠杆：master 刷新的 pnpm store 缓存（只恢复，与 ci.yml 同键）、Wine 供给（apt 安装、Windows Node 下载、`wineboot`）与 `pnpm install` 并发运行、两个阻断表面并发运行——与 `run-gates` 在原生 Windows 上给它们的形状相同——以及按 runner 镜像为键的 apt 归档缓存，使 Wine 的包下载每个镜像版本只付一次。
2026-07-27 实测：热缓存 pull request 运行端到端 2 分 46 秒（准备与缓存恢复约 17 秒，并发安装+供给 33 秒，并发门禁 110 秒），对照 Linux CI 作业的 1.5–2.5 分钟与付费 Windows 通道的 7–9 分钟；冷缓存约多付一分钟。8 核基准腿从未离开队列——受限的 `dsh-ubuntu-*` 池在兄弟 KVM 实验中也被观察到无限排队——因此标准 runner 的数字即为结果，达标不需要更大的机器。
这刻意是一次保真度探针，而非直接替换：Wine 在大小写敏感的 ext4 之上重实现 Win32 API（默认不模拟 NTFS 的大小写不敏感）、不提供 ConPTY、并用自己的安全描述符与 `MoveFileExW` 语义替代——恰是本仓库 `win32.ts` 模块与 PTY 后端关心的表面。实验度量哪些阻断门禁通过、哪些因 Wine 原因而非产品原因失败，以及相对已记录 Windows 基准通道的墙钟成本。 这刻意是一次保真度探针，而非直接替换：Wine 在大小写敏感的 ext4 之上重实现 Win32 API（默认不模拟 NTFS 的大小写不敏感）、不提供 ConPTY、并用自己的安全描述符与 `MoveFileExW` 语义替代——恰是本仓库 `win32.ts` 模块与 PTY 后端关心的表面。实验度量哪些阻断门禁通过、哪些因 Wine 原因而非产品原因失败，以及相对已记录 Windows 基准通道的墙钟成本。
若结论为正则晋升：把 Wine 通道并入为 pull request 的阻断门禁 Windows 信号，将真实 Windows 池降级为 master 串行参照；否则在此记录失败类别并保留该池。 若结论为正则晋升：把 Wine 通道并入为 pull request 的阻断门禁 Windows 信号，将真实 Windows 池降级为 master 串行参照；否则在此记录失败类别并保留该池。