[ChristopherPHolder]

Rebased on the Nx 24 upgrade (#291).

Why

The nx-plugin generators were only covered by virtual-Tree unit tests (createTreeWithEmptyWorkspace), which call the generator factory directly. That left a blind spot around everything the Nx CLI does for real:

• whether generators.json / executors.json / schema.json survive the build's assets globs
• schema.json validation, which only the CLI enforces
• whether the plugin resolves by package name in a consuming workspace
• whether a generated target actually runs afterwards

There was a live bug sitting in that gap, and the e2e found it — see The bug this found below.

What

Adds e2e/nx-plugin-e2e, generated with @nx/plugin:e2e-project. It installs the plugin from the local registry by dist-tag into a real Nx workspace and runs both generators against it.

Local registry scripts

tools/scripts/{start,stop}-local-registry.ts are replaced with the canonical Nx versions. Nx flags the old copy itself when you run the generator:

Your start-local-registry.ts script may be outdated. To ensure that newly generated packages are published appropriately when running end to end tests, update this script to use Nx Release.

The drift cost two things:

	canonical	old copy
teardown	global.stopLocalRegistry = await startLocalRegistry(…)	return value discarded → teardown was a no-op, verdaccio leaked after every run
publishing	releaseVersion + releasePublish	nx run-many --targets publish — only cli has that target, so nx-plugin was never published

The second row is what blocked plugin e2e entirely.

Guards on top of the canonical script

All found by running it:

1. gitPush: false. nx.json sets release.git.push: true for the real release pipeline, and the canonical script only overrides gitCommit/gitTag/stageChanges. The first run printed NX Pushing to git remote "origin" and only failed because the branch had no upstream. With tracking configured it would have pushed.
2. Manifest restore in a finally. releaseVersion writes 0.0.0-e2e into packages/*/package.json on disk, leaving the tree dirty.
3. Explicit registry on releasePublish. targetDefaults["nx-release-publish"] pins the public registry, and that target option beats the npm_config_registry env var startLocalRegistry sets. Without it, publishing tried to authenticate against npmjs and the failure surfaced much later as an opaque ETARGET on install.
4. Fail fast on publish errors. releasePublish resolves with a per-project status map rather than throwing, so a failed publish otherwise limped on.
5. A timeout on registry startup. startLocalRegistry resolves only once verdaccio prints its URL and never rejects if that line doesn't arrive.

Wiring on Nx 24

• @nx/jest:jest no longer exists as an executor, so the e2e target is inferred via a scoped @nx/jest/plugin entry with targetName: "e2e", matching how packages/nx-plugin gets its test target
• the generator also added a workspace-wide @nx/jest/plugin entry with targetName: "jest:test"; dropped it, since the repo already scopes jest inference deliberately
• e2e/nx-plugin-e2e/**/* added to the @nx/eslint/plugin include list
• jest.config.cts, CJS, matching the rest of the workspace after the migration
• new e2e CI job with timeout-minutes: 20, since run-many -t test would not pick the target up

The bug this found

updateDependencies built its dependency map with a bare PLUGIN_NAME as an object key:

const devDeps = {
  '@push-based/user-flow': '^0.19.0',
  PLUGIN_NAME: '^0.0.0',
  '@nx/devkit': '^16.0.0'
};

So nx g @push-based/user-flow-nx-plugin:init wrote a literal "PLUGIN_NAME": "^0.0.0" into the consumer's package.json — a name no registry can resolve, at a version that matched nothing even if it could. A workspace ended up without the plugin it had just been told to install.

Fixed with a computed key, pinned at the version that did the generating rather than a hardcoded ^0.0.0. The assertion for it had been sitting commented out as // TODO implement version test; it is live now, next to one that fails if the literal key returns.

Determinism

Two problems made the suites pass or fail for reasons unrelated to the code.

The target workspace resolved a fresh dependency tree on every run. create-nx-workspace generates a manifest that floats — typescript: ~6.0.3, @types/node: ^24.0.0, prettier: ^3.8.1, tslib: ^2.3.0 — and ships no lockfile, so an upstream release could turn the e2e red with nothing changed here. That scaffold is now frozen into e2e/nx-plugin-e2e/fixtures/workspace/, lockfile included, and the suite npm cis from it. The tree is pinned by integrity hash, and ci additionally refuses to run if the manifest and lockfile drift apart.

Refresh it with npm run e2e:update-workspace-fixture after an Nx bump. The suite asserts the fixture's Nx matches the workspace's and fails with that command in the message, so a forgotten refresh can't silently leave the plugin under test against a stale Nx.

Side effect: the target runs in ~50s rather than 1m14s, since copying a fixture and installing a lockfile beats downloading a template and resolving a tree.

cli-e2e asserted on Node's own output. Eleven assertions required the CLI's stderr to be exactly empty, but Node writes its deprecation and experimental warnings to that same stream — 9 of 23 tests failed on Node 26 with DEP0205 while CI's Node 22 stayed green. The runtime's own lines are now stripped before the comparison, so the tests still catch the CLI writing to stderr without breaking on every new Node release.

Checkout hygiene

• cli-e2e installed the published CLI with npm install -D, rewriting package.json and package-lock.json and only undoing it in teardown — interrupting a run left the checkout carrying a 0.0.0-e2e dependency and a churned lockfile. It installs with --no-save now: node_modules still gets the package, so the tests keep resolving a bare user-flow off the PATH, but both manifests stay byte-identical.
• npm is pinned to the workspace root rather than inheriting the runner's cwd, the registry is stopped when setup throws (vitest skips teardown in that case), and uninstall happens before the registry goes down instead of after.
• .gitignore now ignores nested node_modules; vitest caches into <project>/node_modules and the root-anchored pattern left a test run dirtying the tree.
• The test workspace is created under mkdtemp outside the repo, unique per call, so concurrent worktrees don't collide.

What the e2e asserts

init — registers the plugin under its own name (the regression test for the literal "PLUGIN_NAME"), pins it at the installed version, adds the CLI the executor shells out to, and sets the target generator's default target name in nx.json.

target — writes the rc file, points the generated target at the plugin's executor, and stays resolvable through nx show project, which exercises the plugin's executors.json rather than just re-reading the json we wrote. It also refuses to overwrite an existing target and rejects a run that omits the required projectName.

Not covered: running nx run <project>:user-flow for real. That would install puppeteer into the test workspace and drive Chrome, reintroducing exactly the network and version nondeterminism the fixture lockfile removes.

Verification

Green locally and on CI:

suite	local (Node 26)	CI (Node 22)
cli unit	74/74	pass
nx-plugin unit	13/13	pass
cli-e2e	23/23	pass
nx-plugin-e2e	10/10	pass

cli-e2e was 9 red locally before the stderr change. The plugin e2e installs the same 284 packages on both machines, which is the determinism actually showing up rather than being asserted.

Two failures worth recording, since both were invisible rather than loud:

• A 6-hour CI job. The suite passed in ~40s, then Jest did not exit one second after the test run has completed. Teardown's childProcess.kill() reaches the forked nx wrapper, but verdaccio is its grandchild and survives on Linux holding the pipe. Nx buffers task output until completion, so a passing-but-unterminated task logged nothing at all. Fixed with forceExit, plus --output-style=stream and a job timeout so the next one reports instead of burning six hours.
• Everything proxied through verdaccio. startLocalRegistry sets npm_config_registry for the whole process, so scaffolding dragged ~220MB across 159 scopes through the local proxy. Only @push-based/* needs it; the rest is sent back to the public registry.

[ChristopherPHolder]

Pushed a fix for e2e runtime. The first CI attempt ran for 25 minutes without finishing, so I cancelled it.

Cause: startLocalRegistry sets npm_config_registry for the whole process, so create-nx-workspace scaffolded the test workspace's entire dependency tree through verdaccio. Measured on a local run:

verdaccio storage: 224M across 159 scopes

Every Babel, Jest, SWC and Nx package pulled through a single-threaded Node proxy that also writes each one to disk. Warm locally, stone cold in CI.

Fix: scope the local registry to the packages actually under test rather than raising the timeout.

• create-nx-workspace and the plugin install run with npm_config_registry pointed back at the public registry
• the generated workspace gets an .npmrc with @push-based:registry=<verdaccio> and its auth token

Result, measured cold (rm -rf tmp/local-registry first):

	before	after
verdaccio storage	224 MB	516 KB
scopes proxied	159	1
wall clock (cold, local)	—	39s

$ ls tmp/local-registry/storage/@push-based/
user-flow  user-flow-nx-plugin

Exactly the two packages that need it; everything else resolves from npmjs at full speed. 3 tasks, 3 succeeded, and the working tree comes back clean with versions restored.

[github-actions]

❗❗❗ report generated by this PR ❗❗❗

☕ Order Coffee ☕

Date/Time: 2026-10-24 19:27

Step Name	Gather Mode	Performance	Accessibility	Best Practices	Seo	Agentic Browsing
🧭 Navigate to coffee cart	navigation	92	93	100	83	50
☕ Select coffee	timespan	17/17	-	8/8	-	-
✔ Coffee selected	snapshot	Ø 1/1	16/18	4/4	3/5	1/3
🛒 Checkout order	timespan	19/19	-	8/8	-	-
🧾 Order checked out	snapshot	Ø 1/1	19/20	4/4	3/5	1/3
💌 Submit order	timespan	18/18	-	8/8	-	-
📧 Order submitted	snapshot	Ø 1/1	20/22	4/4	3/5	1/3
🧭 Navigate to github	navigation	98	92	100	83	50

[github-actions]

☕ Order Coffee ☕

Date/Time: 2026-10-24 19:27

Step Name	Gather Mode	Performance	Accessibility	Best Practices	Seo	Agentic Browsing
🧭 Navigate to coffee cart	navigation	93	93	100	83	50
☕ Select coffee	timespan	17/17	-	8/8	-	-
✔ Coffee selected	snapshot	Ø 1/1	16/18	4/4	3/5	1/3
🛒 Checkout order	timespan	19/19	-	8/8	-	-
🧾 Order checked out	snapshot	Ø 1/1	19/20	4/4	3/5	1/3
💌 Submit order	timespan	19/19	-	8/8	-	-
📧 Order submitted	snapshot	Ø 1/1	20/22	4/4	3/5	1/3
🧭 Navigate to github	navigation	96	92	100	83	50