Summary

benchmarks.vlt.sh tracks pacquet (now 0.2.9 ≈ current main) against pnpm 11.3.0 across 8 install variations × 6 fixtures. Two of three original sub-issues landed (#11899, #11901) and one was closed as a debunked diagnosis (#11900). All three retargeted investigations are now fixed on main and waiting for vlt's nightly to pick up the next pacquet release:

• babylon workspaces monorepo DNF → diagnosed as a feature gap, fixed by fix(pacquet/resolving-npm-resolver): honor linkWorkspacePackages for bare-semver deps #11930 (linkWorkspacePackages for bare-semver deps). Residual babylon × {node_modules, lockfile+node_modules, cache+*} DNF cells were a separate panic, fixed by fix(pacquet): port pnpm's workspace-link short-circuit and depPath helpers #11944.
• lockfile+node_modules (every fixture) + cache+lockfile+node_modules large no-op miss → diagnosed as a virtualStoreDir round-trip bug + the missing abbreviated-modified verifier shortcut, fixed by perf(pacquet): unlock no-op short-circuit + port abbreviated-modified verifier shortcut #11931.
• astro deep-transitive tree → diagnosed as a correctness bug (registry optionalDependencies + peerDependenciesMeta silently dropped during deserialization → every optional peer auto-installed → 1535 vs 377 resolutions). Fixed by fix(pacquet/registry): deserialize optionalDependencies and peerDependenciesMeta #11934. Was not any of the four perf candidates listed below; the 4–5× gap was almost all wasted work, not slow work.

Updated chart (2026-05-25)

pacquet 0.2.9 vs pnpm 11.3.0, ratio = pacquet ÷ pnpm. Bold = pacquet slower or DNF.

Numbers from benchmarks.vlt.sh/latest/chart-data.json (commit 572a02be63). babylon cells are flagged pacquet_dnf=true in every variation. These numbers predate #11930, #11931, #11934, and #11944 — the babylon column, the lockfile+node_modules row, and the astro column are all expected to collapse once vlt's nightly picks up the next pacquet release.

Status of original sub-issues

• perf(pacquet): no-op short-circuit when node_modules and lockfile are already consistent #11899 — no-op short-circuit when node_modules and lockfile are already consistent. Landed. The short-circuit existed but never fired on the default macOS/Linux store setup because of the virtualStoreDir round-trip bug fixed by perf(pacquet): unlock no-op short-circuit + port abbreviated-modified verifier shortcut #11931; the lockfile+node_modules cells should now collapse.
• ❌ perf(pacquet): batch/prefetch packument fetches during fresh-resolve tree walk #11900 — batch/prefetch packument fetches during fresh-resolve tree walk. Closed not_planned (context). The prescription was tried in perf(pacquet): prefetch packuments during fresh-resolve tree walk #11903 and regressed the integrated-bench fresh-install path 34-40% — cooperative try_join_all over siblings already gives the same concurrency shape pnpm gets from Promise.all. The astro gap is real but the bottleneck lives elsewhere (likely async_recursion Box-pinning, per-node mutex acquires, or manifest serde — needs profiling).
• pacquet: resolve all workspace projects in the fresh-resolve install path #11901 — resolve all workspace projects in the fresh-resolve install path. Landed. The babylon DNF that surfaced afterwards was a separate linkWorkspacePackages feature gap, fixed by fix(pacquet/resolving-npm-resolver): honor linkWorkspacePackages for bare-semver deps #11930.

Other trackers

• pacquet: pacquet install slower than pnpm install due to syscall contention #11851 — pacquet install slower than pnpm install due to syscall contention. Now the home for install-phase sys-time work that perf(pacquet): close the remaining install-phase sys-time gap to pnpm #11857 used to track.
• perf(pacquet): close the remaining install-phase sys-time gap to pnpm #11857 — closed not_planned (context).
• Rust Roadmap #11633 — Rust Roadmap (Stage 2 resolver port covers the deep-tree work).

Follow-ups from local vlt verification

Local re-bench against pacquet 0.2.10 (comment, investigation) confirmed the astro column collapse but surfaced two residual gaps. Both are tracked as follow-up issues; this issue stays open until they ship and vlt's nightly confirms.

• bug(pacquet/lockfile): PkgNameVerPeer::without_peer panics with MismatchParenthesis on some snapshot keys #11939 — bug(pacquet/lockfile): PkgNameVerPeer::without_peer panics with MismatchParenthesis on some snapshot keys. Cause of the 37–38× ratios in babylon × node_modules / lockfile+node_modules. Fixed by fix(pacquet): port pnpm's workspace-link short-circuit and depPath helpers #11944 — root cause was an unported piece of pnpm's resolver: workspace link: nodes weren't short-circuited (depth === -1 arm in resolvePeersOfNode) and ended up flowing through peer resolution, producing depPaths of the shape link:<rel-path>(<peers>) that downstream code mis-parsed. PR also closed three smaller divergences: pkgIdWithPatchHash construction at four call sites was using the wrong strip semantics; is_runtime_dep_path and try_get_package_id from deps/path weren't ported.
• perf(pacquet/install): no-op short-circuit fires after verify_lockfile_resolutions, defeating the fast path on cache wipes #11940 — perf(pacquet/install): no-op short-circuit fires after verify_lockfile_resolutions, defeating the fast path on cache wipes. Cause of the lockfile+node_modules row being 1.6–7.2× across fixtures even though the on-disk state is already a no-op. Two ways forward — reorder so the no-op fires first, or port pnpm's earlier shortcut.

Retargeted work

1. • Babylon DNF on every variation → pacquet: linkWorkspacePackages ignored for bare-semver deps (workspace packages bypassed; can wrong-install or 404) #11929 → fixed by fix(pacquet/resolving-npm-resolver): honor linkWorkspacePackages for bare-semver deps #11930 (pacquet/resolving-npm-resolver: honor linkWorkspacePackages for bare-semver deps). Diagnosis: long-standing feature gap, not a recent regression. Babylon's root depends on @dev/build-tools: ^1.0.0 (a workspace package); pacquet's resolver only routed workspace:-prefixed specs through the workspace path, so the bare semver 404'd against npm. Residual babylon × {node_modules, lockfile+node_modules, cache+*} DNFs after fix(pacquet/resolving-npm-resolver): honor linkWorkspacePackages for bare-semver deps #11930 were a separate without_peer panic, fixed by fix(pacquet): port pnpm's workspace-link short-circuit and depPath helpers #11944.
2. • Astro deep-tree perf → fixed by fix(pacquet/registry): deserialize optionalDependencies and peerDependenciesMeta #11934 (pacquet/registry: deserialize optionalDependencies and peerDependenciesMeta). Diagnosis was not in the candidate list — PackageVersion (the type the npm resolver deserializes registry version metadata into) was missing those two fields, so serde_json::to_value(picked) produced a manifest with neither. Downstream, extract_children saw no optionalDependencies edges (astro's sharp got dropped) and extract_peer_dependencies saw no peerDependenciesMeta (every optional peer became "required" and the autoInstallPeers fallback in hoist_peers installed it). On astro that cascaded unstorage's 19 optional peers (@azure/*, @vercel/*, @netlify/blobs, @upstash/redis, ...) plus all their transitives into the install: 1535 resolutions vs pnpm's 377. After fix(pacquet/registry): deserialize optionalDependencies and peerDependenciesMeta #11934 the lockfile shape matches pnpm (377 resolutions, astro@5.18.1(rollup@4.60.4)(typescript@5.9.3) root suffix, sharp present); warm-store install 39.6 s → 8.5 s (4.7× faster, pnpm 7.0 s); warm-cache hyperfine pacquet 670 ms vs pnpm 1.27 s (1.89× faster). See comment for full numbers.
3. • lockfile+node_modules no-op miss → fixed by perf(pacquet): unlock no-op short-circuit + port abbreviated-modified verifier shortcut #11931 (pacquet: unlock no-op short-circuit + port abbreviated-modified verifier shortcut). Two bugs in one cell: modules-yaml's resolve_virtual_store_dir joined the stored relative virtualStoreDir with modules_dir without normalising .., so the on-disk path never byte-matched the config side and the perf(pacquet): no-op short-circuit when node_modules and lockfile are already consistent #11899 short-circuit silently skipped every install with a global store. Even after that, the resolution verifier paid a full-meta GET per name because pacquet only ported 2 of pnpm's 4 publish-timestamp layers; the abbreviated-modified shortcut + on-disk mirror are now in place. Cold-cache svelte ran 2.16 s → 0.71 s (3.0× faster, vs pnpm 0.54 s); the residual ~150 ms is install-bootstrap overhead.

Methodology note

vlt invokes pacquet as pacquet install (no flags), so every variation that wipes the lockfile lands on the fresh-resolve path — currently the least-developed install path (Stage 2 in #11633). pacquet's integrated-benchmark mostly exercises --frozen-lockfile, which is why internal numbers look better than the published vlt chart. Reproduce with:

BENCH_INSTALL_PNPM="corepack pnpm@latest install --ignore-scripts --silent"
BENCH_INSTALL_PACQUET="pacquet install"

Goal

Pacquet ≥ pnpm on every (variation × fixture) cell in the vlt.sh chart. The astro column collapsed cleanly under pacquet 0.2.10. The remaining gap reduces to one follow-up: #11940 (no-op short-circuit ordering). Once that lands and vlt's nightly picks up the resulting release, the chart should show pacquet ahead of pnpm on every cell that previously regressed.

Written by an agent (Claude Code, claude-opus-4-7).