Dev Wip Package Refer

dev-wip-package-refer

The "consume a WIP upstream package via a sibling-relative file: dep" pattern. Lets a consumer project depend on an in-development library/framework without publishing each iteration to npm.

For the inverse — editing the upstream itself — see dev-wip-package-upstream-wt-dev.

When to reach for this pattern

Use when the upstream library is itself in active development by the same maintainer (or a tightly coordinated team) and the iteration loop matters: edit upstream → consumer sees the change immediately on next pnpm install, no publish step. Trades multi-machine setup pain for fast iteration.

Don't use it when the upstream is stable and shipped to npm — a normal versioned dep is simpler. Also don't use it for adversarial / arms-length dependencies — file: paths assume both repos live on disk under your control.

How it resolves

The consumer's package.json declares the dep with a path-style spec:

{
  "dependencies": {
    "@org/pkg":    "file:../upstream/packages/pkg",
    "@org/pkg-rt": "file:../upstream/packages/pkg-runtime"
  }
}

pnpm/npm resolve file:../upstream/... against the consumer project root, so the upstream sibling is always at <consumer-root>/../upstream/. Identical on every machine as long as the sibling layout is preserved — clone the consumer at $HOME/repos/foo/consumer, the upstream at $HOME/repos/foo/upstream, and file:../upstream just works. No env-specific config.

Real example (zudo-doc):

"@takazudo/zfb":                     "file:../zfb/packages/zfb"
"@takazudo/zfb-adapter-cloudflare":  "file:../zfb/packages/zfb-adapter-cloudflare"
"@takazudo/zfb-runtime":             "file:../zfb/packages/zfb-runtime"
"@takazudo/zudo-design-token-panel": "file:../zdtp/packages/zudo-design-token-panel"

→ resolves to ../zfb/packages/zfb and ../zdtp/packages/zudo-design-token-panel.

What the consumer needs in place

For pnpm install to succeed in the consumer, the upstream sibling must:

1. Exist on disk at ../upstream/ (clone before installing).
2. Be at a known SHA so the consumer's behavior is reproducible — see "Pinning" below.
3. Have any required build artifacts present. pnpm hard-copies file: deps at install time. If the dep is "source + a built binary" (e.g. a Rust CLI shipped via the npm package), the binary must be built before the consumer installs. If the dep is "TypeScript + a dist/" the dist must be built first.

If any of those are missing, the consumer's pnpm install either succeeds with broken/stale state or fails at a postinstall hook.

Pinning — single source of truth for the SHA

The consumer pins which upstream SHA it depends on. Two places to pin:

1. In CI workflow env vars — e.g. ZFB_PINNED_SHA, ZDTP_PINNED_SHA declared at the workflow env: level of every workflow that runs pnpm install. CI clones the upstream at that SHA before installing. This is the source of truth.
2. (Optional) A framework-pins.json that both CI workflows and a local bootstrap script read from. Reduces "edit 3 YAML files per bump" to "edit 1 JSON file." Use when you have more than ~2 workflow files.

Bumping the pin = the consumer adopts the upstream change. The change is reviewed and tested in CI's clean clone.

CI side — clone-then-install

The consumer's CI workflow must clone the upstream at the pinned SHA into ../upstream/ before pnpm install runs in the consumer. Shape:

env:
  UPSTREAM_PINNED_SHA: <full SHA>

jobs:
  build:
    steps:
      - name: Checkout consumer
        uses: actions/checkout@v5

      - name: Clone pinned upstream sibling
        run: |
          git clone https://github.com/<org>/<upstream-repo>.git ../upstream
          git -C ../upstream checkout "$UPSTREAM_PINNED_SHA"

      # OPTIONAL: build upstream artifacts here if the npm package needs them
      # (Rust binary, dist/, etc.). For heavy builds, do it in a separate
      # job and upload the artifact, then download it into ../upstream/ in
      # each consumer job. See zudo-doc's `build-zfb` job for that pattern.

      - name: Setup Node + pnpm
        # ...

      - name: Install consumer deps
        run: pnpm install

For expensive upstream builds (Rust toolchain, large bundles), split into a dedicated build job that uploads the binary as an artifact; consumer jobs cp it into ../upstream/target/release/ before pnpm install. See zudo-doc's .github/workflows/pr-checks.yml build-zfb job for a reference shape.

Local side — bootstrap script for multi-machine setup

The brittle part of this pattern is "new machine = clone two repos in the right layout + build their artifacts before pnpm install." Solve it with a bootstrap script that does what CI does:

pnpm setup:upstream     # ensure all WIP-sibling upstreams are present and built

The script:

1. Reads the pinned SHA(s) from the single source of truth (CI workflow env vars or framework-pins.json).
2. For each upstream:
   • If ../upstream/ doesn't exist → git clone <url> ../upstream, then git checkout <SHA>.
   • If ../upstream/ exists with a clean tree → git fetch origin && git checkout <SHA> (matches CI).
   • If ../upstream/ exists with a dirty tree → refuse to touch it (you have in-flight upstream edits; resolve first or pass --force-checkout).
3. Build upstream artifacts if missing (Rust binary, dist/, etc.). Skip if cache is fresh.
4. Run pnpm install in the consumer.

The "refuse on dirty tree" rule is what prevents stomping on a parallel upstream-edit session — see dev-wip-package-upstream-wt-dev for why the upstream root is shared.

A starter Node.js skeleton (adjust to the consumer's needs):

// scripts/setup-upstream.mjs
import { execSync } from "node:child_process";
import { existsSync, readFileSync } from "node:fs";
import { resolve } from "node:path";

const projectRoot = process.cwd();
const pkg = JSON.parse(readFileSync(resolve(projectRoot, "package.json"), "utf8"));

// Discover sibling-relative file: deps and group by upstream root.
const upstreams = new Map(); // siblingDir -> { siblingPath }
for (const [name, spec] of Object.entries(pkg.dependencies ?? {})) {
  if (typeof spec !== "string" || !spec.startsWith("file:..")) continue;
  const siblingDir = spec.slice("file:".length).split("/")[0]; // "../upstream"
  const siblingPath = resolve(projectRoot, siblingDir);
  if (!upstreams.has(siblingDir)) upstreams.set(siblingDir, { siblingPath });
}

// For each upstream, load (sha, gitUrl, buildSteps) from `framework-pins.json`
// (or grep the workflow YAML) and ensure-clone-build.
// ...left to the consumer to fill in based on its specific upstreams.

Wire the script into package.json as a regular script ("setup:upstream": "node scripts/setup-upstream.mjs"). Don't run it from postinstall — that would loop. Document it as the first thing to run on a new machine.

Comparison with alternatives

The file: pattern wins when iteration speed on upstream matters and the team can absorb the bootstrap step. The published-package pattern wins when the upstream stabilizes.

Bumping the pin

Edit the SHA in the single source of truth (CI workflow env var or framework-pins.json). Commit + push the consumer-side change. CI re-clones the upstream at the new SHA in its clean checkout and validates. Locally re-run pnpm setup:upstream to mirror the new pin if you want to test before pushing — otherwise trust CI.

For the how of preparing the new upstream SHA itself (an upstream PR, merge, watch CI green), see dev-wip-package-upstream-wt-dev.

Anti-patterns

• Don't git checkout random branches on ../upstream/. That checkout is shared with every consumer using file:../upstream/... and with any concurrent Claude session. Use a worktree — see dev-wip-package-upstream-wt-dev.
• Don't run the bootstrap from postinstall. Postinstall already runs on every pnpm install — the bootstrap script itself runs pnpm install, so wiring it as postinstall loops.
• Don't commit absolute paths to package.json. file:../upstream/... is portable; file:/home/you/repos/upstream/... is not.
• Don't pin the SHA only in pnpm-lock.yaml. file: deps point at on-disk paths, not SHAs — the lockfile is stable across pin bumps and won't help reproducibility. The SHA must live in CI env vars (or framework-pins.json).