fix(repo): repair the vm runner and config subpath after 11.0.0 by systemfsoftware-maker · Pull Request #110 · systemfsoftware/stryker-js-effect

systemfsoftware-maker · 2026-09-25T07:10:41Z

Summary

Six regressions users hit after upgrading to 11.0.0 now work again: defineConfig/mergeConfig config files, strict-pnpm installs of the default vm runner, and four vm runner divergences from vitest run.

import { defineConfig, mergeConfig } from '@systemfsoftware/stryker-js/config' loads and typechecks again, and const config: StrykerConfig = defineConfig({...}) typechecks. The 11.0.0 StrykerConfig class statics are removed; the changeset is a major and gives the migration.
A pnpm project that installs only stryker-js and vitest completes a vm dry run. The harness is now located from its own install instead of from the project's cwd. effect is now a required peer rather than an optional one: optional peers are never installed, so the CLI failed to read its config without effect, even though it imports effect at runtime.
import.meta.vitest inside a string, template, comment or regex literal is no longer rewritten, so RuleTester-style fixtures pass.
A test file with no tests passes when that project sets passWithNoTests.
Setup files and helpers outside the project root get the harness vitest, so a module-scope beforeEach no longer fails with "Vitest failed to find the runner".
A failed dry run names each failing test and its message, both in the error and at the default log level.

CI repairs

main's CI was already red (typescript-checker#lint:tsgo, instrumenter and test-contribution timeouts), and this branch surfaced three more flakes. All are fixed at the root, with no suppressions and no raised timeouts:

TypeScriptCompiler is declared as a Context.Service class, which clears the lint:tsgo error.
The concurrency opt-in selection scenario instruments only the fixtures that carry each mutator (1.4 s to 0.3 s), and the test-contribution laws draw only the fields the judgement reads (125 s to 25 s under CI settings). Both had crossed the 30 s timeout under CI coverage.
The vm runner concurrency scenario proves overlap with a rendezvous instead of a 300 ms window.
The worker-options round-trip laws no longer draw object keys with JSON escapes. V8 13.6 corrupts those keys once a lone-backslash key has been parsed in the process (nodejs/node#63785), which falsified a correct codec.
The passWithNoTests harness test resolves its sandbox through realPath, because macOS links /tmp to /private/tmp.
The vm parity differential records the unmutated baseline once per fixture instead of once per mutant (565 s to about 140 s).

Not in this PR

#100's enterprise-fixture parity half. Its hand-written live-export and Vite-id patches were dropped in favour of the vm runner redesign, #111, which is layer 2 of stack #112 on top of this PR. The four enterprise e2e configs stay on the vitest runner until then.

Fixes #104
Fixes #105
Fixes #106
Fixes #107
Fixes #108
Related: #100

Validation

Local: pnpm format:check, pnpm typecheck, pnpm test, pnpm check:ci and the changeset check pass. CI is green on ubuntu and macOS, including the e2e lane. For #105 I packed stryker-js and installed it into a fresh pnpm project (default isolated linker) whose only direct dependencies were stryker-js and vitest. There, a vm dry run and a mutation run both exit 0. That strict-pnpm install has no in-repo test: the e2e bake installs every tarball with npm.

The vm session added a failed `(no test suite)` result for every discovered file that registered no tests, ignoring Vitest's `passWithNoTests`. The flag now travels from the resolved Vitest project config through the schema, defaults and extraction, and the session consults the file's own project before reporting the failure, mirroring Vitest's `!file.tasks.length && !config.passWithNoTests` check Closes #107

The vm runner handed its in-memory vitest only to importers inside the sandbox prefix or to files under node_modules/<served name>. A setup file or helper outside the project root therefore received the real, runner-less vitest: a module-scope registration either failed with "Vitest failed to find the runner" or silently registered nothing, stopping the dry run as init-failed. Track the modules a run loads — any importer inside an active sandbox, any importer carrying a run salt, and any module reached from one of those — and serve the harness vitest to all of them. The served shims now read the current run state when their registration API is called, so an unsalted, shared module registers against the current test file's runner rather than a stale session's Closes #108

The vm runner replaced every textual `import.meta.vitest` in a loaded module with a global regex, so a string or template literal, a comment, or a regex literal that mentioned the marker reached the program corrupted. A suite whose fixtures contain the marker (a lint-rule RuleTester, for example) failed on the vm runner while passing under `vitest run`, because Vitest only rewrites modules it discovered through `includeSource`. Replace the regex with a scanner that rewrites only real member expressions and copies string, template, comment and regex text byte for byte. The scanner tracks template substitutions (an expression inside `${…}` is still rewritten), escaped quotes and backticks, and picks between a regex literal and division from the last significant token. Both the host-transform path (`hostResult`) and the loaded-source path of the transform plugin use the scanner, so in-source `if (import.meta.vitest)` tests still register and run Closes #106

… cwd The vm test runner's Vitest host worker resolved @systemfsoftware/stryker-vm-harness/package.json with a createRequire rooted at process.cwd(). Under pnpm's default isolated node_modules layout the harness is installed beside @systemfsoftware/stryker-js and is not visible from the project root, so a project that installed only stryker-js and vitest failed its dry run with: Cannot find module '@systemfsoftware/stryker-vm-harness/package.json'. The lookup now uses createRequire(import.meta.url), resolving the harness from where the harness itself is installed — the same self-resolution worker-client.ts already uses — while the project's own vitest keeps resolving from the project. The dist-less fallback also pointed at a stale src/shell/vitest-host directory; it now points at src/vitest-host, so a source checkout with no build still starts the host worker Closes #105

The compound-pack refactor moved the config helpers onto a `StrykerConfig` class, so `@systemfsoftware/stryker-js/config` stopped exporting `defineConfig` and `mergeConfig` and `StrykerConfig` became a class-instance type. A `stryker.config.ts` written against the 10.1.1 surface then failed at load time, and `const config: StrykerConfig = defineConfig({ … })` failed to typecheck, with nothing in the 11.0.0 changelog explaining why. `defineConfig` returns the four 10.1.1 overloads (object, promise, and the `(env: ConfigEnv) => …` factory forms) and `mergeConfig(defaults, overrides)` composes presets again, both as named exports of `./config`. `StrykerConfig` names the partial options a config file writes. The schema-consuming runtime values moved out of `stryker-config.schema.ts` into `config/default-options.ts` so the schema module only declares schemas and types Closes #104

A failing dry run reported only counts, so the stage failure, its cause and the machine-mode error envelope named neither the failing tests nor their assertion messages. The dry-run decision now carries every failing test's name and failure message (DryRunFailed.failedTests, fed from the runner's failed TestResults); the stage failure's reason renders them and the dry run logs them at error level, so the default-level log and the error event both name each failure Closes #100

The real-vitest reference arm re-copied the fixture and spawned a cold 'vitest run --update' for every mutant, and both arms ran the full vm mutation engine to list the same mutants. Record the unmutated baseline once per fixture, copy it per mutant, and share one cached engine report between the arms; the mocking mutant case drops from 65s to 32s locally Refs #110

effect-tsgo rejects a Context.Service assigned to a variable; a named shape interface keeps the declaration emit free of the handle's private slot type Refs #110

The selection scenario instrumented the whole 528-mutant corpus eight times and crossed the 30s timeout under CI coverage; it now reads one placement site per mutator and keeps the same checks Refs #110

…ment reads Generating whole mutation reports per draw pushed three laws past the 30s timeout on CI Refs #110

A fixed 300ms window flaked under CI load; each suite now waits for its sibling to start Refs #110

…he bug V8 13.6 corrupts escaped JSON object keys once a lone-backslash key is parsed (nodejs/node#63785); the law domain now draws only unescaped keys and the ineffective seed pin is removed Refs #110

macOS links /tmp to /private/tmp, so results keyed by the real path never matched the sandbox path the test looked up Refs #110

effect was an optional peer although the CLI imports it at runtime, so a pnpm project that installed only stryker-js and vitest failed to read its config. Required peers are installed by pnpm and npm, so the #105 install works and Effect users still share one copy Refs #105

systemfsoftware-maker added 13 commits September 25, 2026 06:00

refactor(repo): declare the TypeScript compiler service as a class

337a2f9

effect-tsgo rejects a Context.Service assigned to a variable; a named shape interface keeps the declaration emit free of the handle's private slot type Refs #110

test(repo): select concurrency mutators over the fixtures that hold them

5ff5f1a

The selection scenario instrumented the whole 528-mutant corpus eight times and crossed the 30s timeout under CI coverage; it now reads one placement site per mutator and keeps the same checks Refs #110

test(repo): draw test-contribution commands from the fields the judge…

d548671

…ment reads Generating whole mutation reports per draw pushed three laws past the 30s timeout on CI Refs #110

test(repo): prove concurrent vm runners overlap with a rendezvous

da0259d

A fixed 300ms window flaked under CI load; each suite now waits for its sibling to start Refs #110

test(repo): keep worker-options law keys clear of the V8 JSON key-cac…

e700b9f

…he bug V8 13.6 corrupts escaped JSON object keys once a lone-backslash key is parsed (nodejs/node#63785); the law domain now draws only unescaped keys and the ineffective seed pin is removed Refs #110

test(repo): resolve the pass-with-no-tests sandbox to its real path

4fc29f1

macOS links /tmp to /private/tmp, so results keyed by the real path never matched the sandbox path the test looked up Refs #110

blacksmith-sh Bot mentioned this pull request Sep 25, 2026

refactor(repo): run testRunner vm on the vitest runner and delete the vm harness #111

Merged

systemfsoftware-maker added this pull request to stack #112 September 25, 2026 14:15

ryanleecode approved these changes Sep 25, 2026

View reviewed changes

ryanleecode merged commit e37907b into main Sep 25, 2026
10 checks passed

ryanleecode deleted the bug-fixes2 branch September 25, 2026 14:19

Navigation Menu

fix(repo): repair the vm runner and config subpath after 11.0.0 - #110

fix(repo): repair the vm runner and config subpath after 11.0.0#110
ryanleecode merged 14 commits into
mainfrom
bug-fixes2

systemfsoftware-maker commented Sep 25, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

systemfsoftware-maker commented Sep 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

CI repairs

Not in this PR

Validation

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

systemfsoftware-maker commented Sep 25, 2026 •

edited

Loading