fix(core): speed up location boot and VCS routes on Windows - #60569

Issue for this PR

No issue: found while profiling the server on Windows.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

On Windows each git process costs ~35-40 ms, and opening a Location ran several of them back to back while the agent, model, provider and integration lists waited. Plugin activation awaits every state's notify, and two of them do git work: the VCS branch refresh (2 sequential rounds for a clone, 3 without a remote) and the location watcher's repository probe, a rev-parse that project resolution had already answered. That was ~175 ms of the ~200 ms activation flush on reopen.

• plugin/vcs/git.ts: the default branch reads the remote list together with origin's HEAD, and reads local branches with init.defaultBranch in one more round only when no remote HEAD answers. vcs.status no longer checks HEAD first (diffing an unborn HEAD lists nothing). vcs.diff fetches the batched patch in the same round as the change list, since chunks name their own files, and branch mode reads HEAD and the default branch together.
• git.ts: repo.discover is cached by the .git metadata fingerprint, as Project.resolve already is, so the location watcher and snapshots reuse project discovery.
• util/which.ts: the Windows lookup stats once per PATHEXT extension instead of once per case and does not throw on each miss: 14.6 → 6.2 ms for a missing command, which the Azure plugin looks up on every Location boot. A test checks it against which with awkward PATH and PATHEXT values.
• boot-work.test.ts: a step spawns 8 git processes instead of 9, because snapshot discovery is reused.
• packages/server/script/benchmark-server.ts (new): boots opencode serve from any checkout with isolated state and replays a client opening a location: open, each endpoint warm, reopen, and /api/info latency while review diffs run, with git spawns per phase. --compare prints before/after.

Windows 11, Ryzen 7 7800X3D, Git for Windows 2.48.1. Median of 8 server launches on a generated 3,000-file repo shaped like a clone, with two edits and a new file:

On effect-smol (2,163 files, clean tree): reopen 245.9 → 156.8 ms, vcs.status 88.5 → 50.4 ms, branch diff 247.8 → 178.7 ms. The working diff is flat there (93.6 → 94.5 ms): on a clean tree the batched patch is a wasted parallel process. Starting four git processes at once also makes the worst single event-loop stall during a diff ~4 ms longer; p99 is unchanged.

Not included: starting mingw64\bin\git.exe instead of the cmd\git.exe launcher saves another ~10 ms per git call, but the synchronous part of CreateProcess grows with the image (~3 ms for the 45 KiB launcher, ~8.4 ms for the 4 MiB git.exe) and lands on the server thread. With it, /api/info p99 during diffs went from 1.0 to 9.1 ms.

How did you verify your code works?

• core, via bun run script/test.ts: boot-work, vcs, git, snapshot, project, worktree, vcs-init, provider-azure, location-watcher-policy, which, shell, formatter and ripgrep (217 pass). server: vcs, worktree, session-diff and process tests.
• bun run check, and bun run typecheck:tsc in core and server.
• From packages/server, against a dev worktree: bun run script/benchmark-server.ts --source <dev worktree> --runs 8 --json before.json, then bun run script/benchmark-server.ts --runs 8 --compare before.json. Also with --directory on effect-smol.

Visual

<svg xmlns="http://www.w3.org/2000/svg" width="240" height="72" viewBox="0 0 240 72" fill="none">
  <style>
    text{font:500 9px system-ui,-apple-system,"Segoe UI",sans-serif;fill:currentColor;opacity:.55}
    .k{fill:currentColor;opacity:.55}
    .ok{stroke:currentColor;stroke-width:1.6;stroke-linecap:round;stroke-linejoin:round;opacity:.8}
    .cur{fill:currentColor;opacity:0}
    @media (prefers-reduced-motion:no-preference){
      .r1,.a{animation:r1 3.6s linear infinite}
      .r2{animation:r2 3.6s linear infinite}
      .r3{animation:r3 3.6s linear infinite}
      .oka{animation:oka 3.6s linear infinite}
      .okb{animation:okb 3.6s linear infinite}
      .cur{animation:cur 3.6s linear infinite}
    }
    @keyframes r1{0%{opacity:.12}1%,28%{opacity:.95}30%,100%{opacity:.55}}
    @keyframes r2{0%,28%{opacity:.12}30%,57%{opacity:.95}59%,100%{opacity:.55}}
    @keyframes r3{0%,57%{opacity:.12}59%,86%{opacity:.95}88%,100%{opacity:.55}}
    @keyframes oka{0%,28%{opacity:0}31%,100%{opacity:.8}}
    @keyframes okb{0%,86%{opacity:0}89%,100%{opacity:.8}}
    @keyframes cur{0%{opacity:.4;transform:translateX(0)}88%{opacity:.4;transform:translateX(156px)}92%,100%{opacity:0;transform:translateX(156px)}}
  </style>
  <text x="6" y="23">before</text>
  <text x="6" y="54">after</text>
  <rect class="k r1" x="46" y="13" width="50" height="13" rx="3"/>
  <rect class="k r2" x="99" y="13" width="50" height="13" rx="3"/>
  <rect class="k r3" x="152" y="13" width="50" height="13" rx="3"/>
  <path class="ok okb" d="M209 19.5l3 3 6-6.5"/>
  <rect class="k a" x="46" y="43" width="50" height="3.4" rx="1.7"/>
  <rect class="k a" x="46" y="47.8" width="50" height="3.4" rx="1.7"/>
  <rect class="k a" x="46" y="52.6" width="50" height="3.4" rx="1.7"/>
  <path class="ok oka" d="M103 50l3 3 6-6.5"/>
  <rect class="cur" x="45" y="8" width="1.5" height="54" rx=".75"/>
</svg>

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR