fix(ci): stop compiling better-sqlite3 from source on Node 24

[benjaminshafii]

The failure

OpenWork Tests on Linux dies during pnpm install --frozen-lockfile (e.g. #3371, run 30575171143):

.../better-sqlite3 install: prebuild-install warn install No prebuilt binaries found (target=24.13.0 runtime=node arch=x64 libc= platform=linux)
.../better-sqlite3 install: gyp http GET https://nodejs.org/download/release/v24.13.0/node-v24.13.0-headers.tar.gz
.../better-sqlite3 install: gyp ERR! stack AggregateError [ETIMEDOUT]:
[ELIFECYCLE] Command failed with exit code 1

Root cause

better-sqlite3@11.10.0 publishes prebuilds for node-v108/115/127/131 — no node-v137, which is Node 24. Every CI job runs node-version: 24, so prebuild-install always missed, fell back to node-gyp rebuild, and reached out to nodejs.org for headers mid-install. Any hiccup on that request fails the whole install. It was never green-by-design, just green-when-the-network-cooperated.

Fix

• Bump better-sqlite3 to ^12.11.1 in apps/server and apps/desktop. That release ships node-v127/137/141/147 and electron-v133 (we're on Electron 35) prebuilds, so installs download a binary instead of compiling.
  • 12.0.0 was a build-matrix-only major (dropped EOL Node 18 / Electron 26-28) — no JS API change, and our usage is new Database() + prepare() in apps/server/src/opencode-db.ts.
  • Not 13.x: it just landed (2026-07-29) and switches to N-API, which also trips minimumReleaseAge: 4320.
• Cache the pnpm store in ci-tests.yml (same pattern already used by alpha-macos-aarch64.yml / release-macos-aarch64.yml) and install with --prefer-offline. The log showed reused 0 on every progress line — CI was re-downloading all 1453 packages each run, which is both slow and more network to flake on.

Verification

Run locally on the dev base in a fresh worktree (macOS arm64):

Command	Result
pnpm install --frozen-lockfile	better-sqlite3 install: Done in seconds, build/Release/better_sqlite3.node present with no obj.target compile artifacts — prebuild path, no node-gyp
node -e "new (require('better-sqlite3'))(':memory:')..."	sqlite 3.53.2 | node 25.2.1
pnpm --filter openwork-server test	600 pass, 10 skip, 0 fail
pnpm --filter @openwork/desktop test	174 pass, 1 skip, 0 fail
pnpm check:outbound-access	pass (35 hosts covered)
pnpm --filter @openwork/desktop typecheck:electron	pass
pnpm --filter @openwork/app test	566 pass / 2 fail — pre-existing on dev, unrelated: message-list-loading.test.tsx fails only inside the full-suite run (target.addEventListener is not a function from a DOM global clobbered by another file) and passes in isolation; dev's own CI run is green on this lane

The real proof for this change is this PR's own OpenWork Tests run: the install step should no longer print No prebuilt binaries found / gyp output at all.

No user-visible surface changes (dependency + workflow only), so no fraimz was recorded — per the agreement in the request, CI green plus the local runs above is the evidence.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openwork-app	Ready	Preview	Jul 31, 2026 10:41am
openwork-den	Ready	Preview	Jul 31, 2026 10:41am
openwork-den-worker-proxy	Ready	Preview	Jul 31, 2026 10:41am
openwork-diagnostics	Ready	Preview	Jul 31, 2026 10:41am
openwork-landing	Ready	Preview, v0	Jul 31, 2026 10:41am

[benjaminshafii]

CI evidence

Run 1 — the compile is gone (job 91135979688):

.../better-sqlite3 install$ prebuild-install || node-gyp rebuild --release
.../better-sqlite3 install: Done          # 0.3s, no "No prebuilt binaries found", no gyp output

Run 2 — the store cache is doing its job (job 91137548193):

Cache restored from key: Linux-pnpm-1f5fed3e...
Progress: resolved 1459, reused 1451, downloaded 0, added 1459, done     # was: reused 0, downloaded 1451
Progress: resolved 50, reused 50, downloaded 0, added 50, done           # evals workspace

better-sqlite3 doesn't even run its install script on the cached run.

All checks green on both runs, including openwork-tests (blacksmith-4vcpu-ubuntu-2204) and openwork-tests (macos-14) — which also confirms the 2 @openwork/app failures I saw locally are a local-environment artifact, not this change.