Effect 4

Apply Effect 4 semantics from project-local architecture, the narrowest relevant reference, and source matching the target's installed packages.

Workflow

Do not activate this workflow merely because a file imports effect. For nontrivial Effect work:

1. Resolve the target package or workspace and its exact installed effect and relevant @effect/* versions. If effect is not 4.x, stop because this skill does not apply. For 3.x, point the user to the v3 skill: npx skills add PaulRBerg/agent-skills#shelved --skill effect-ts-v3.
2. Confirm lockstep versions: every @effect/* package must match the installed effect version exactly. Report a mismatch instead of coding around it.
3. Inspect neighboring services, layers, errors, schemas, runtime boundaries, and tests. Local conventions decide organization. Installed package evidence decides API facts.
4. Read references/critical-rules.md, then only the task-specific references below.
5. Verify every uncertain import, signature, or behavior against the package installation visible to the target workspace before editing. Do not port v3 names from memory. Many were renamed or removed.
6. Implement the smallest pattern consistent with the project and run the narrowest test or typecheck covering the changed semantics.

Evidence Order

Use the target workspace's manifest and lockfile to identify versions. Prefer, in order:

1. the installed package's src/, README, tests, and changelog.
2. its emitted declarations when source is not shipped.
3. the matching official package artifact or source tag.

For monorepo source, tests, or cross-package search, run scripts/effect-source.sh. It keeps a cached clone of Effect-TS/effect checked out at the newest release, fetching at most once per 24 hours and otherwise reusing the cache, and prints path= and release=. When release differs from the installed effect version, read the matching tag without checking it out: git -C <path> grep <pattern> effect@<version> -- packages/ or git -C <path> show effect@<version>:<file>. Never edit the cache.

Do not install or update dependencies solely to obtain documentation. Do not trust v3 examples or source that does not match the target's installed version. If exact behavior cannot be verified, stop rather than guessing.

Former @effect/platform, @effect/rpc, @effect/cli, @effect/cluster, and related modules now ship inside effect under subpaths such as effect/http, effect/http-api, effect/rpc, effect/cli, effect/ai, effect/sql, effect/schema, effect/reactivity, and effect/workflow. Check the installed package.json exports for the exact set. These modules are documented @stability unstable and may break in minor releases. There are no effect/unstable/* compatibility paths. Separate packages remain for platforms, drivers, and providers: @effect/platform-*, @effect/sql-*, @effect/ai-*, @effect/atom-*, @effect/opentelemetry, and @effect/vitest. Their modules are also @stability unstable.

From 4.0.2, every module and directly importable export declares @stability. Check the tag before you depend on an API. stable follows strict semver. experimental modules, such as effect/graphql and parts of effect/net, may break in patch releases. Some root modules are also unstable, such as Arbitrary, Crypto, FileSystem, and Path.

Reference Router

For effect/http, effect/http-api, effect/rpc, effect/cli, collection utilities, deprecations, or constructor lookup, inspect the installed package source directly instead of loading a local API inventory.

Boundaries

• Keep pure helpers, constants, and path manipulation pure unless an Effect boundary provides a concrete dependency, testability, resource-safety, or error-model benefit.
• Preserve existing domain facades and service/runtime boundaries unless the user requested redesign.
• At IO boundaries, prefer typed failures and scoped resources. Choose Schema-backed errors/models only when encoding or boundary validation is needed.
• Do not broaden environment requirements merely to replace a small platform call.

For changes, completion requires code consistent with local Effect architecture, selected references and installed source where needed, and the narrowest test/typecheck that exercises the changed semantics. Read-only work requires evidence for the reported conclusion. Finish with ### ⚡ Effect — ✅ change complete after verified edits or ### ⚡ Effect — 🔎 reviewed, no files written for read-only work, one sentence naming the boundary or pattern used, and ### 🧪 Verification with exact scoped commands/results. If required validation is incomplete, use ### ⚡ Effect — ⛔ blocked instead. Add ### ⚠️ Limitation only for non-blocking caveats. Never decorate typed errors, Schema messages, logs, tests, generated JSON/API responses, or command output.