[systemfsoftware-maker]

Summary

A consumer can now import any service contract without pulling in its implementation. No *.service.ts file that declares a Context.Service tag builds or exports a Layer any more. Each of those Layers now lives in a driver module (src/drivers/<what-it-binds>.ts) in the owning package, and the e2e harness follows the same pattern. This is layer L1 of the plan in docs/plans/2026-10-09-2237-refactor-ports-split-capability-packages-plan.md. L2 splits packages/stryker-js into one package per capability, and it depends on this change.

• ADR-0003 (docs/adr/0003-service-ports-declare-no-layer.md) supersedes ADR-0001's *.service.ts row and extends its src/drivers/ row. It was drafted as 0002, but refactor(stryker-js): retire the Checker plans and protocol modules #260 took that number on main first, so it is now 0003 and cites ADR-0001's rows by name rather than by line.
• Gate: gritlint.json enables the cell-architecture pack, and flake.lock pins the root systemfsoftware flake input, which builds gritlint, to upstream main ba6664c18 (systemfsoftware#701 then #704). The two 45f2deb entries that arrive through the published stryker-js release's own lock are unchanged. pnpm lint:conventions in check:ci now refuses a service module that builds, re-exports, or aliases a Layer.
• attw joins check:ci (pnpm attw, through each package's existing attw --pack .). @systemfsoftware/oxlint-ignorer-config was the one published package without the script. It now has one, plus the ESM-only .attw.json that the other ESM packages carry. The first CI run on this PR then showed that stryker-js-plugin-interface and stryker-js-cli-contract had the script but no .attw.json, so they fail NoResolution and CJSResolvesToESM. They get the same file, and a local turbo run attw --continue now passes all 51 tasks.
• Packs: .compound-engineering/config.yaml declares all four: cell-architecture, boundary-testing, schema-laws, and package-topology.
• Breaking changes: static layer members and the *Live aliases are gone. The replacements are Engine.stage/Engine.forStream (previously RunEnvironment.stage/forStream), RunEvent.drainLayer, Worker.workerOptionsLayer/workerTelemetryLayer, and the ./runtime entry checkerRuntimeLayer(options). Changesets: core-service-ports-declare-no-layer, plugin-runtime-worker-layers, and typescript-checker-runtime-layer are major; vitest-runner-session-driver, oxlint-ignorer-config-attw, and plugin-interface-attw-config are none.

How an agent consumes this

• gritlint finding. gritlint check --format json emits findings[] with rule (cell-architecture/service-exports-no-layer, a stable reason code), path, line, and message, plus a summary with per-rule file counts. The exit code is 1 on a finding and 0 when clean.
  • The rule is multifile. It reports once per directory, naming the alphabetically first *.service.ts file at line 1, so an agent has to search that directory's service modules for the Layer. The message's trailing (directory ...) names the directory, and the message says so.
  • The message ends with the next action, verbatim from the pinned pack: Next: move the Layer to src/drivers/<what-it-binds>.ts as layer, delete its export from the service module, and import the driver at the composition root.
  • The rule refuses a Layer.* call, a class field named layer or *Layer, a *Live export, a re-export of layer/*Layer/*Live, and an exported binding (const, let, var, default, including annotated, as, and satisfies forms) bound to one. A type-only re-export is allowed. export * from and a Layer under an unrelated name are not reported; the pack README lists these gaps.
• attw step in check:ci.
  • pnpm attw runs turbo's attw task, which runs attw --pack . in each published package.
  • The @systemfsoftware/arethetypeswrong-cli fork prints one JSON line per package: status, packageName, packageVersion, problems[] (each with kind, entrypoint, and resolutionKind, where kind is attw's stable problem code such as NoResolution), and problemCounts. It then prints a one-line hint naming the --include flags that add more detail.
  • A failing package exits non-zero, and check:ci records s=1 and keeps running the remaining steps.
  • The fix for NoResolution on an ESM-only package is the .attw.json shown in the attw output section below. The envelope carries no schema-version field; that is the fork's format, and this PR does not change it.
• ADR-0003's rule.
  • The *.service.ts row ("Never holds: a Layer value, a function returning a Layer, a static layer or *Layer member, a *Live name; in a file that declares a tag, driver imports") is enforced by the gritlint rule id above, except for the driver-import clause.
  • The driver-import clause is checked by review. Its check is the command in the ADR's Confirmation section, git grep -nE "from '(\./|\.\./)+drivers/" -- '*.service.ts', whose expected output is only plugin-loader.service.ts, a module with no tag that plan U11/U12 rehomes. In L2, tsc -b on the contracts package takes over.
• CI. This PR adds or changes no CI job. check:ci gains the pnpm attw step inside the existing job, so there is no new $GITHUB_STEP_SUMMARY or annotation surface.

Review

• Run: report-only ce-code-review, full spine, at 5d8a97ee4. Reviewers: correctness, adversarial, project-standards, testing, maintainability, api-contract, and learnings, plus a validator.
• Coverage gap: the cross-model review did not run, because no other-model CLI (codex, grok, cursor-agent, opencode) is installed on the host. The adversarial lens ran in-process on the same model as the other reviewers, so no finding here has independent cross-model corroboration.
• Verdict: Ready with fixes. R6-R8 and R11 belong to L2, as the plan says.

#	Sev	Finding	Disposition
1	P2	stryker-js-plugin-runtime/README.md named the removed Worker.WorkerOptions.layer	Fixed: it now names Worker.workerOptionsLayer. A repo-wide sweep for removed names found one more, docs/solutions/runtime-errors/stage-scope-closed-with-exit-void-hides-run-outcome.md, which now names Engine.stage and temporaryDirectoryLayer.
2	P2	The gate misses a re-exported or aliased Layer	Fixed upstream in systemfsoftware#704 (merged as ba6664c18) and re-locked here in 0bb9c503e. #704 also refuses the annotated, as, satisfies, and export default forms. With the re-lock, the clean tree still exits 0, so no *.service.ts here re-exports or aliases a Layer. The probes below show each form refused.
3	P2	attw skipped the published @systemfsoftware/oxlint-ignorer-config	Fixed: the package now has the attw script, the fork devDependency (catalog, one pnpm-lock.yaml importer entry), and .attw.json. Output is below.
4/5	P3	ADR's driver-import clause was unenforced and broken by plugin-loader.service.ts	Fixed in the ADR: the clause now binds only files that declare a tag, which is true today (grep below). The tagless plugin-loader.service.ts is named as moving in plan U11/U12, and Confirmation names the check that takes over.
6	P3	JSDoc on VitestSessionInput and VitestSessionShape was dropped in the move	Fixed: both blocks were restored verbatim from main.

Driver imports in *.service.ts files (the grep for #4/#5):

$ git grep -nE "from '(\./|\.\./)+drivers/" -- '*.service.ts'
packages/stryker-js/src/plugin-loader.service.ts:17:import { importModule } from './drivers/config.js'

plugin-loader.service.ts declares no tag.

Evidence

attw on @systemfsoftware/oxlint-ignorer-config

Without .attw.json:

{"status":"ok","packageName":"@systemfsoftware/oxlint-ignorer-config","packageVersion":"0.1.2","types":{"kind":"included"},"problems":[{"kind":"NoResolution","entrypoint":".","resolutionKind":"node10"},{"kind":"NoResolution","entrypoint":".","resolutionKind":"node16-cjs"}],"problemCounts":{"NoResolution":2}}
[ELIFECYCLE] Command failed with exit code 1.

With .attw.json, byte-identical to the 14 sibling ESM packages:

$ attw --pack .
{"status":"ok","packageName":"@systemfsoftware/oxlint-ignorer-config","packageVersion":"0.1.2","types":{"kind":"included"},"problems":[],"problemCounts":{}}
exit 0

pnpm lint:conventions probes

These runs use gritlint built from the locked input ba6664c18 (/nix/store/wmsl4wda…-gritlint-0.1.0). Each probe file was packages/stryker-js/src/zz-probe.service.ts, deleted after its run, and the tree was clean afterwards.

Tree	Exit	Output
Clean head	0	cell-architecture/service-exports-no-layer: evaluated 39 file(s); clean: 1119 file(s) selected
Probe export { layer as gitDiffLayer } from './drivers/git-diff.js'	1	packages/stryker-js/src/Sandbox.service.ts:1: cell-architecture/service-exports-no-layer (the directory's first service module); evaluated 40 files
Probe import { layer } from './drivers/git-diff.js' plus export const gitDiff = layer	1	Same finding
Probe export { layer as GitDiffLive } from './drivers/git-diff.js'	1	Same finding
Probe deleted	0	evaluated 39 file(s), clean

With the 6a7c247b3 lock that came before, the first two probes exited 0. The gritlint check --format json . finding for the re-export probe:

{"rule":"cell-architecture/service-exports-no-layer","path":"packages/stryker-js/src/Sandbox.service.ts","line":1,"message":"A `*.service.ts` module in this directory builds or exports a `Layer`: a\n`Layer.*` call, a class field named `layer` or `*Layer`, a `*Live` export, a\nre-export of `layer`, `*Layer` or `*Live`, or an export bound to one. The\nfinding names the directory's first service module; search every `*.service.ts`\nin the directory. Next: move the Layer to `src/drivers/<what-it-binds>.ts` as\n`layer`, delete its export from the service module, and import the driver at\nthe composition root. (directory `packages/stryker-js/src`)"}

Local verification

Verification was targeted because the host has little memory.

• Merged head (origin/main merged up, ADR renumbered):
  • tsc -b in packages/stryker-js passes.
  • The core unit and property suite passes: 191 files, 1,137 tests.
  • turbo build for @systemfsoftware/stryker-js passes, as does api:check.
  • lint:conventions exits 0.
• Earlier on this branch:
  • tsc -b and oxlint pass for each changed package.
  • attw passes for each changed package.
  • Plugin-runtime, typescript-checker, and vitest-runner tests passed, with 6, 191, and 23 tests respectively.
  • After the comments-only JSDoc restore, vitest-runner was rerun: 5 of 6 files passed. kill-first-ordering.integration.test.ts failed to start because a vitest pool worker timed out ([vitest-pool-runner]: Timeout waiting for worker to respond). That is host memory pressure, not an assertion failure. CI runs the file.
• Mutation: no local mutation run; that happens in CI on main.

Merge order

This PR shares 8 files with #258 (stryker/mutant-quality-l1-reasons, also based on main): packages/stryker-js/etc/stryker-js.api.md, src/Serve/Serve.cell.ts, src/mutation-reporting.service.ts, and five integration tests and fixtures. Whichever PR lands second will need origin/main merged in, and the API report regenerated.

[ryanleecode]

Verified: CI green on 0bb9c50, diff read, no open threads.

[ryanleecode]

Closing without merging, superseded: this moves Layers into src/drivers/.ts, a technical-kind folder, which the cell-architecture placement settled in systemfsoftware#708 and #728 rejects. The rule now: a pure implementation stays as static layer on the service class, a driver-backed implementation sits in a flat sibling module beside its service that only composition roots and tests import, and there are no drivers/ folders. It's also 106 files on a day-old base. The rework comes as a fresh PR off main. The branch stream-a/l1-ports-topology is kept.