# @systemfsoftware/oxlint-plugin-effect-workflow ## 5.1.3 ### Patch Changes - Bump the stryker catalog to pull in the latest patch releases of the stryker-js ecosystem. ## 5.1.2 ### Patch Changes - Dependencies resolve to Effect 4.0.0-rc.116 and Vitest 5. ## 5.1.1 ### Patch Changes - `workflow-variant-constructed` no longer reports a workflow whose decision annotation names a local bound to a constructed value, such as `Result.Result` where `const accepted = new Admitted({})`. That name is a value rather than a variant class the file declares, and the annotation writes it where the compiler expects a type, so the report no longer speaks for it. A variant class the file declares with no construction site still fails the run: construct it with `new X(…)` or `X.make(…)`, or remove it from the union and from every dispatch over it. ## 5.1.0 ### Minor Changes - `make-file-location` and `workflow-file-make-presence` now recognize `Workflow.total` and `Workflow.andThen` as constructors alongside `Workflow.make`. A workflow-named file that builds its decision with a total decider or a chained pair is no longer reported for a missing `Workflow.make` call, and both rules' messages name all three constructors instead of `Workflow.make` alone. The location rule's one-construction-per-file limit counts totals and composites too, so a workflow-named file that opens a second construction is now reported. Move any construction the rules newly report out of the file that opens it and into the workflow-named file that owns it. - `workflow-variant-constructed` is a new recommended rule at `error`. A workflow whose decider declares a decision or error variant that no code path in the same file constructs now fails lint, instead of shipping a variant a consumer of the union can never receive. The declared channels come from the decider's `Result.Result<…, …>` return annotation — written inline, through an alias declared in the file, or as a `S.Union([…])` value the annotation names. A variant counts as constructed when the file builds it with `new X(…)` or `X.make(…)` anywhere, not only inside the decider. A variant whose class is imported rather than declared in that file is not judged, and type-test files are exempt. Construct the variant in the file, or delete it from the union and from every dispatch over it. ## 5.0.1 ### Patch Changes - Releases the workspace so its published versions track the shared dependency graph this change moves. ## 5.0.0 ### Major Changes - The core regime keys on the `Workflow.make` boundary; the complement gains a complexity ceiling. - `workflow-match-exhaustive` no longer reads the filename: the gate is the make callee boundary (import binding + member `make` + argument containment, module-scope references followed, shadow-correct). Identical dispatch outside a make body produces no diagnostic. - New `make-body-purity`: references inside make bodies resolve only to parameters, const locals, and audited-pure imports; control flow is banned with the one first-statement converging guard; unclassifiable references report honestly as unresolvable rather than passing. Test files are exempt — fixtures exercise decisions without the production regime binding them. - New `no-domain-branching-density` in core: per-function McCabe CC outside make bodies, ceiling 17 — the lowest measured value the tree passes with zero waivers (max measured 17; the 15 functions over 10 are the recorded extraction backlog, not retrofit targets). - Both make-boundary rules fix classifier defects their first workspace run exposed: builtin globals with empty defs take the named-global triage, and `as const` type subtrees leave the value-reference walk. ### Patch Changes - New version is published through npm trusted publishing, so it carries a provenance attestation you can verify. - Peer Effect requirement advances to 4.0.0-rc.112. No API changes. ## 4.2.0 ### Minor Changes - Two new recommended rules run against every workflow file. `damp-workflow-stem` requires the file's name to be a hyphenated phrase of two to five lowercase words whose camelCase form equals the file's single export, so a file named for its capability bucket or a bare noun now fails. `workflow-file-make-presence` refuses a workflow file that never constructs its decision with `Workflow.make`. Rename failing workflow files to the decision they make and align each export name, or move a file that owns no decision next to its caller without the workflow suffix. ## 4.1.0 ### Minor Changes - Recommended now reports extra non-schema value exports and re-exports from a workflow file. A single-segment workflow module may publish exactly one non-schema value. Schema declarations and type-only exports stay. Every re-export form is an error, including exporting an imported binding. If the new diagnostic fires, delete the extra export or import the name from the module that declares it. ## 4.0.2 ### Patch Changes - Re-published against the oxc-based instrumenter: workspace dependency ranges move to the new instrumenter major; no package's own behavior changes in this release. ## 4.0.1 ### Patch Changes - Peer Effect requirement advances to 4.0.0-rc.112. No API changes. ## 4.0.0 ### Major Changes - New rule `make-command-schema`, enabled as an error in the recommended config: the command position of `Workflow.make` may not hold a value laundered into looking like a schema class. Upgrading with the recommended config on will fail builds that currently pass. It reports three shapes, all of which a type checker accepts: a type assertion, a non-null assertion, or a `satisfies` clause at the command position; a value assembled by `Object.assign`, `Object.create`, `Reflect.construct` or a `Proxy` wrapper; and a binding that exists only as a `declare`. It stays silent wherever the type checker already refuses — a plain class, an object literal, a struct schema, a primitive — so it adds no second report on code that already fails to compile. It is also silent on a factory call that returns the class, and on these constructors reached through an alias rather than written out. ### Patch Changes - The peer requirements for `effect` and for the Effect test-runner integration now accept any compatible `4.0.0-rc` release, instead of demanding one exact release candidate. Installing alongside a newer release candidate no longer reports an unmet peer dependency or resolves a second copy of `effect` into the dependency tree. ## 3.0.0 ### Major Changes - `make-body-purity` reports every import referenced inside a `Workflow.make` body, except the sealed pure `effect` surface. A decision is the innermost point of the sandwich, so imports run toward it and never out of it: the reader imports the workflow, and nothing sits beneath the pure core. A body reaching a sibling module invents a layer there whose purity nothing checks — this rule visits make bodies only, so it never reads the module the body reached. The rule previously exempted eight relative specifiers, which is why such a body could pass. The exemption keyed on the filename an author typed, so adding a line certified a module without reading it, renaming a file un-certified one whose contents had not changed, and a specifier like `./Survivors.js` certified that name in any package in any directory. Expect new findings in any package whose decision bodies call helpers from neighbouring modules. Each has two resolutions: move the referenced code into the deciding file, or move the decision into the file that already holds the code. One decision, one file. Passing the helper in as a parameter is not a third option — a pure helper does not earn a requirement. Where the reference is a decoder, the decode belongs at the edge and its result passes into the decision as data. - The core regime keys on the `Workflow.make` boundary; the complement gains a complexity ceiling. - `workflow-match-exhaustive` no longer reads the filename: the gate is the make callee boundary (import binding + member `make` + argument containment, module-scope references followed, shadow-correct). Identical dispatch outside a make body produces no diagnostic. - New `make-body-purity`: references inside make bodies resolve only to parameters, const locals, and audited-pure imports; control flow is banned with the one first-statement converging guard; unclassifiable references report honestly as unresolvable rather than passing. Test files are exempt — fixtures exercise decisions without the production regime binding them. - New `no-domain-branching-density` in core: per-function McCabe CC outside make bodies, ceiling 17 — the lowest measured value the tree passes with zero waivers (max measured 17; the 15 functions over 10 are the recorded extraction backlog, not retrofit targets). - Both make-boundary rules fix classifier defects their first workspace run exposed: builtin globals with empty defs take the named-global triage, and `as const` type subtrees leave the value-reference walk. - Thirteen `workflow-*` rules are removed, and with them two AST helpers that had no other consumer. `effect-workflow` now ships five rules; `effect-dmmf` re-exports the smaller set. Removed: `workflow-single-function-export`, `workflow-command-object`, `workflow-declaration-form`, `workflow-schema-required`, `workflow-either-inhabited`, `workflow-typeid-required`, `workflow-typeid-shared-per-union`, `workflow-union-schema-declared`, `workflow-no-unconstructed-variant`, `workflow-no-throw`, `workflow-no-async`, `workflow-single-path`, `workflow-no-ambient-impurity`. Kept: `workflow-no-panic-vocabulary`, `workflow-match-exhaustive`, `workflow-no-effect-import` and `workflow-property-test-shape` in `configs.recommended`, plus `workflow-inline-schemas` registered but not recommended. Every removal names what refuses the violation instead. Three are refused by `Workflow.make`'s own constructor types, which reject a workflow whose error channel is `never` (`__WORKFLOW_ERROR_CHANNEL_IS_NEVER__`), one whose decision channel is (`__WORKFLOW_DECISION_CHANNEL_IS_NEVER__`), and an error carrying no `_tag` to dispatch on (`__WORKFLOW_ERROR_CHANNEL_CARRIES_NO_TAG__`) — the diagnostic arrives at the construction site, names its own fix, and travels to consumers through the type rather than through an installed lint config. Five more are compiler diagnostics already at `error` in `packages/tsconfig/effect.json`: `asyncFunction` covers `workflow-no-async`, and `globalFetchInEffect`, `globalDateInEffect`, `globalRandomInEffect` and `processEnvInEffect` cover `workflow-no-ambient-impurity`. What survives as a lint rule is what neither mechanism can see: an import edge, an identifier's vocabulary, and one rule over hand-authored test files. The remaining removals are deliberately unenforced rather than re-homed. `workflow-no-throw` and the declaration-shape rules asserted properties no type refuses and no diagnostic reports, and each was keyed on a filename, so it never fired on the violation it existed to catch. `docs/2026-08-15-orphaned-cell-constraints.md` records every one with what owning it would take. Consumers spreading `configs.recommended` need no change; a config naming a removed rule by key must drop that key. ### Minor Changes - Three new rules, a retargeted test-placement taxonomy, and one removal. `make-file-location` allows a workflow constructor only in the workflow module that owns it, at most once per module. `schema-declaration-location` requires a schema declaration to live in a schema module, or the workflow module that owns it. A binding whose initializer returns something other than a schema — a type guard, a decoder, an encoder, an arbitrary — is a use and is not reported. `test-placement` narrows which tests may sit beside source, requires every other test to live in the package test directory, and adds `tests-dir-helpers-in-fixtures`. It also removes `in-source-test-targets-private`, which `effect-dmmf` no longer re-exports — drop the entry if you set it. Each rule reports the exact shape it expects. ### Patch Changes - Remove citations to specs that do not exist Each leaf cited a `skill://architect-` spec, an id minted inside one, or a gate this repo no longer carries. None of those resolve — no `architect-*` skill exists, and the ids lived in an operator-layer file that could never ship with the clone. Two `meta.docs.description` strings in `effect-middleware` are the consumer-visible half of this: they ended in a parenthetical naming a spec the consumer cannot read. Each sentence already stated its constraint in full, so the citation was the only part removed. Where a citation carried content — which constraints a package deliberately leaves to review, and why each is out of mechanical reach — the constraint is now named in place instead of pointed at. - New version is published through npm trusted publishing, so it carries a provenance attestation you can verify. - State the canonical-identifier contract once Six leaves carried the same rule: a rule matches the canonical identifier only, its suite carries a near-miss valid case proving the alias does not fire, and widening the match makes every one of those cases pass vacuously. It now lives once in the hub leaf that governs the whole family, so a leaf below inherits it rather than restating it. One clause was genuinely package-specific and stayed: a computed `Effect['fn']` still counts as an exported `Effect` value under `store-effect-fn-required`, which the shared rule does not say. Also corrects the turbo cycle recorded in two places. Both chains from `effect-executor` to `effect-dmmf` are real — `effect-cell-types` dev-depends on `oxlint-config` directly and again through `effect-gherkin-spec` — so the two were true at different granularity rather than in conflict, and both now say so. The closing edge is absent today; the rules exist to keep it absent. ## 2.0.0 ### Major Changes - `make-body-purity` reports every import referenced inside a `Workflow.make` body, except the sealed pure `effect` surface. A decision is the innermost point of the sandwich, so imports run toward it and never out of it: the reader imports the workflow, and nothing sits beneath the pure core. A body reaching a sibling module invents a layer there whose purity nothing checks — this rule visits make bodies only, so it never reads the module the body reached. The rule previously exempted eight relative specifiers, which is why such a body could pass. The exemption keyed on the filename an author typed, so adding a line certified a module without reading it, renaming a file un-certified one whose contents had not changed, and a specifier like `./Survivors.js` certified that name in any package in any directory. Expect new findings in any package whose decision bodies call helpers from neighbouring modules. Each has two resolutions: move the referenced code into the deciding file, or move the decision into the file that already holds the code. One decision, one file. Passing the helper in as a parameter is not a third option — a pure helper does not earn a requirement. Where the reference is a decoder, the decode belongs at the edge and its result passes into the decision as data. - The core regime keys on the `Workflow.make` boundary; the complement gains a complexity ceiling. - `workflow-match-exhaustive` no longer reads the filename: the gate is the make callee boundary (import binding + member `make` + argument containment, module-scope references followed, shadow-correct). Identical dispatch outside a make body produces no diagnostic. - New `make-body-purity`: references inside make bodies resolve only to parameters, const locals, and audited-pure imports; control flow is banned with the one first-statement converging guard; unclassifiable references report honestly as unresolvable rather than passing. Test files are exempt — fixtures exercise decisions without the production regime binding them. - New `no-domain-branching-density` in core: per-function McCabe CC outside make bodies, ceiling 17 — the lowest measured value the tree passes with zero waivers (max measured 17; the 15 functions over 10 are the recorded extraction backlog, not retrofit targets). - Both make-boundary rules fix classifier defects their first workspace run exposed: builtin globals with empty defs take the named-global triage, and `as const` type subtrees leave the value-reference walk. - Thirteen `workflow-*` rules are removed, and with them two AST helpers that had no other consumer. `effect-workflow` now ships five rules; `effect-dmmf` re-exports the smaller set. Removed: `workflow-single-function-export`, `workflow-command-object`, `workflow-declaration-form`, `workflow-schema-required`, `workflow-either-inhabited`, `workflow-typeid-required`, `workflow-typeid-shared-per-union`, `workflow-union-schema-declared`, `workflow-no-unconstructed-variant`, `workflow-no-throw`, `workflow-no-async`, `workflow-single-path`, `workflow-no-ambient-impurity`. Kept: `workflow-no-panic-vocabulary`, `workflow-match-exhaustive`, `workflow-no-effect-import` and `workflow-property-test-shape` in `configs.recommended`, plus `workflow-inline-schemas` registered but not recommended. Every removal names what refuses the violation instead. Three are refused by `Workflow.make`'s own constructor types, which reject a workflow whose error channel is `never` (`__WORKFLOW_ERROR_CHANNEL_IS_NEVER__`), one whose decision channel is (`__WORKFLOW_DECISION_CHANNEL_IS_NEVER__`), and an error carrying no `_tag` to dispatch on (`__WORKFLOW_ERROR_CHANNEL_CARRIES_NO_TAG__`) — the diagnostic arrives at the construction site, names its own fix, and travels to consumers through the type rather than through an installed lint config. Five more are compiler diagnostics already at `error` in `packages/tsconfig/effect.json`: `asyncFunction` covers `workflow-no-async`, and `globalFetchInEffect`, `globalDateInEffect`, `globalRandomInEffect` and `processEnvInEffect` cover `workflow-no-ambient-impurity`. What survives as a lint rule is what neither mechanism can see: an import edge, an identifier's vocabulary, and one rule over hand-authored test files. The remaining removals are deliberately unenforced rather than re-homed. `workflow-no-throw` and the declaration-shape rules asserted properties no type refuses and no diagnostic reports, and each was keyed on a filename, so it never fired on the violation it existed to catch. `docs/2026-08-15-orphaned-cell-constraints.md` records every one with what owning it would take. Consumers spreading `configs.recommended` need no change; a config naming a removed rule by key must drop that key. ### Minor Changes - Three new rules, a retargeted test-placement taxonomy, and one removal. `make-file-location` allows a workflow constructor only in the workflow module that owns it, at most once per module. `schema-declaration-location` requires a schema declaration to live in a schema module, or the workflow module that owns it. A binding whose initializer returns something other than a schema — a type guard, a decoder, an encoder, an arbitrary — is a use and is not reported. `test-placement` narrows which tests may sit beside source, requires every other test to live in the package test directory, and adds `tests-dir-helpers-in-fixtures`. It also removes `in-source-test-targets-private`, which `effect-dmmf` no longer re-exports — drop the entry if you set it. Each rule reports the exact shape it expects. ### Patch Changes - Remove citations to specs that do not exist Each leaf cited a `skill://architect-` spec, an id minted inside one, or a gate this repo no longer carries. None of those resolve — no `architect-*` skill exists, and the ids lived in an operator-layer file that could never ship with the clone. Two `meta.docs.description` strings in `effect-middleware` are the consumer-visible half of this: they ended in a parenthetical naming a spec the consumer cannot read. Each sentence already stated its constraint in full, so the citation was the only part removed. Where a citation carried content — which constraints a package deliberately leaves to review, and why each is out of mechanical reach — the constraint is now named in place instead of pointed at. - New version is published through npm trusted publishing, so it carries a provenance attestation you can verify. - State the canonical-identifier contract once Six leaves carried the same rule: a rule matches the canonical identifier only, its suite carries a near-miss valid case proving the alias does not fire, and widening the match makes every one of those cases pass vacuously. It now lives once in the hub leaf that governs the whole family, so a leaf below inherits it rather than restating it. One clause was genuinely package-specific and stayed: a computed `Effect['fn']` still counts as an exported `Effect` value under `store-effect-fn-required`, which the shared rule does not say. Also corrects the turbo cycle recorded in two places. Both chains from `effect-executor` to `effect-dmmf` are real — `effect-cell-types` dev-depends on `oxlint-config` directly and again through `effect-gherkin-spec` — so the two were true at different granularity rather than in conflict, and both now say so. The closing edge is absent today; the rules exist to keep it absent.