Update Dependencies

Update npm/pnpm Dependencies

Overview

Pinchy is a pnpm workspace with several independent package.jsons (root, packages/web, each packages/plugins/*, plus the standalone docs/ package). A blind "bump everything to latest" misses two kinds of traps: packages that are pinned together and packages that are pinned apart for a documented reason.

Scope

Excluded from this skill — use update-openclaw instead: openclaw (dev dependency, the pinned core runtime) and openclaw-node (our own client library). OpenClaw upgrades have repeatedly broken Pinchy in ways pnpm test/tsc don't catch (session keys, tools.allow semantics, config.apply timing) — bumping them needs the dedicated release-notes review procedure, not a routine sweep. If pnpm outdated shows a newer openclaw/openclaw-node, list it in your report but do not touch it here; tell the user to run update-openclaw separately.

Procedure

1. Enumerate outdated packages per workspace, not just root:

   pnpm outdated                              # root
   pnpm -C packages/web outdated
   for d in packages/plugins/*/; do pnpm -C "$d" outdated; done
   cd docs && pnpm outdated                    # standalone package.json/lockfile

   If node_modules isn't installed in the current worktree, every row shows "missing (wanted X)" — that's not a signal, ignore it. The wanted → latest columns are what matter.

2. Bucket every outdated package into one of three groups before touching any file:

   • Safe patch/minor bump. Same major version, no known pairing constraint. Bump directly.
   • Paired — must move together, or not at all. Check before bumping:
     • next + eslint-config-next (Next.js ships them in lockstep; eslint-config-next also gates the eslint-v10 blocker below).
     • Any two packages from the same vendor that render/parse together (e.g. an avatar/icon library split into a core + style package) — if only one has a newer major, leave both pinned until the other catches up.
     • Any package listed in pnpm.patchedDependencies (root package.json) — a local patch is pinned to an exact version. Bumping needs its own step: verify the patch still applies after the bump (pnpm install fails loudly if it doesn't), and if it no longer applies, regenerate it and confirm the behavior it exists for still works (check pnpm patch history / memory for why the patch exists before assuming it's obsolete).
   • Known blocker or deliberately deferred — do not bump, state why:
     • eslint 9 → 10 is blocked: eslint-config-next pulls in eslint-plugin-react, whose peer range caps ESLint below 10; forcing it past that crashes pnpm lint at runtime. Recheck when eslint-config-next ships a major that drops this constraint.
     • A dependency's major version just became latest days/weeks ago (check npm view <pkg> time or the dist-tags) — prefer the newest release within the previous major and flag the fresh major as a separate, deliberate follow-up once it has real-world mileage.
     • A dependency that's a hand-rolled integration point, not a thin passthrough (e.g. a PDF/document engine used directly by our own extraction/render code) — a major bump there needs its own PR with the relevant test suite run against real sample data, not a bundled sweep.

3. Apply the safe-bump and paired-bump groups, then verify:

   pnpm install
   pnpm -C packages/web lint
   pnpm test                      # root script fans out to workspaces
   pnpm -C packages/web test:db
   pnpm test:scripts
   pnpm build
   pnpm format
   cd docs && pnpm build           # if docs/ package.json changed

   Run the relevant pnpm -C packages/web test:e2e:<suffix> for any plugin whose external-API client library was bumped (e.g. googleapis → test:e2e:email).

4. Report, don't auto-commit. This touches runtime dependencies, not just app code — present the diff (git status --short, git diff --stat per workspace) and the three buckets (bumped / paired-and-bumped / deliberately-skipped-with-reason) and let the user decide to commit. Give majors you deliberately skipped their own follow-up mention so nobody re-attempts them blind next sweep.