SuperAPI AI Directory

java

Coding & development

View repository

Updated 11-07 08:29

About this skill

Java (language skill)

This is a knowledge skill, not an agent: it loads from context whenever Java is written or reviewed, and the role agents (/be, /rev, /arch) route into it for the language-level rules. Framework depth — Spring, persistence, API design, messaging — stays with /be (../../backend/java/backend-developer/SKILL.md) and its ../../backend/java/backend-developer/references/java-expertise.md; this skill owns the language.

Trigger

Use this skill when:

  • Writing, editing, or reviewing .java files
  • Changing pom.xml, build.gradle, build.gradle.kts, or gradle/libs.versions.toml
  • Diagnosing a javac, Error Prone, or JVM test failure
  • Designing a Java component — domain model, null contract, concurrency topology
  • Planning tests for Java code
  • Adding a dependency to a JVM build

Do NOT load it for:

  • Kotlin — that is /be's kotlin reference
  • Spring/API/persistence how-to — that is /be's java-expertise reference

Context

This skill exists to make Java reviews boring. Illegal states cannot compile, so the reviewer argues about the domain instead of hunting null checks. Null flows are verified by a checker at build time, so "can this be null?" has a mechanical answer. Architecture rules are executable tests, so the layering survives its tenth contributor. Where the language has moved — records, sealed types, virtual threads, ScopedValue — the doctrine moves with it and states which runtime each piece of advice applies to.

Documentation Lookup (MANDATORY)

Before implementing any feature, check current documentation. Java ships a new release every six months, preview APIs mutate between releases, and the tooling stack (Error Prone, NullAway, PIT, ArchUnit) moves independently; recall is not a source.

Context7 MCP

Use Context7 MCP to retrieve up-to-date documentation for any library or tool:

  1. Resolve library: Call mcp__context7__resolve-library-id with the library name
  2. Query docs: Call mcp__context7__query-docs with the resolved library ID and your question

When to use: JUnit 6 APIs, ArchUnit rule syntax, Testcontainers modules, NullAway and Error Prone flags, pitest configuration, jqwik arbitraries.

Example queries:

  • "JUnit 6 parameterized test sources"
  • "ArchUnit 1.4 layeredArchitecture syntax"
  • "Testcontainers Spring Boot ServiceConnection supported containers"
  • "NullAway OnlyNullMarked option"
  • "pitest incremental analysis configuration"
  • "jqwik combinators for recursive records"

Web Research

Use WebSearch and WebFetch for anything version- or advisory-shaped:

Question Source to check
Is a feature final or preview on the deployed JDK the JEP page — its status line, not blog posts
What changed in a JDK release the release notes / JEP index for that exact JDK
Current version of a library its repository / Maven Central, at task time
Advisory against a dependency the OSV database

Java-specific lookup rules

  • Preview APIs are looked up against the current JEP number, never recalled. Preview features rename between releases — structured concurrency's Joiner API changed between the JDK 25 preview (JEP 505) and the JDK 26 preview (JEP 525). Code recalled from an earlier preview round will not compile on the next JDK; find the JEP that ships with the project's JDK and read that one.
  • Framework claims are verified against the deployed major. Spring Framework 7 is JSpecify-annotated and recommends NullAway; Spring 6 used its own null-safety annotations. Null-safety, baseline-JDK, and namespace advice all differ by major — check which one the project actually runs before asserting anything.

Rule: When uncertain about any API, configuration, or best practice — search first, code second.

Versions

Technology Version Notes
Java 25 (LTS) The baseline this skill targets; 26 is non-LTS, 27 expected Sept 2026
ScopedValue final since 25 JEP 506
Stream Gatherers final since 24 JEP 485 — Stream.gather(...)
FFM final since 22 java.lang.foreign — the JNI replacement
Structured concurrency PREVIEW JEP 505 (5th preview, 25) → JEP 525 (6th preview, 26, Joiner renames); expected final in 27
Virtual-thread pinning fix JDK 24 JEP 491 — synchronized no longer pins
JSpecify 1.0 org.jspecify:jspecify:1.0.0; Spring Framework 7 is annotated with it
Error Prone ≥ 2.36 The floor NullAway requires
NullAway 0.12.x OnlyNullMarked mode for incremental adoption
JUnit 6.0.x GA 2025-09; one version across Platform/Jupiter/Vintage; Java 17 baseline
PIT (pitest) 1.19.4+ Incremental analysis; JUnit Platform engines (JUnit 6 Jupiter) via the separate pitest-junit5-plugin (1.2.x)
jqwik 1.10.x Property-based testing
ArchUnit 1.4.x Architecture rules as tests
Testcontainers current With Spring Boot @ServiceConnection
OSV-Scanner current Dependency CVEs in CI
CycloneDX current SBOM plugins for Gradle and Maven

Volatility note. Versions above are current as of Aug 2026 — re-verify before pinning. Java ships every six months and preview JEPs renumber each release; the structured-concurrency row in particular expires when JDK 27 finalizes it.

The Doctrine — the J-standards (BLOCKING at review)

/be's numbered Engineering Standards 1–7 (facts-only Javadoc, no narration comments, naming, builder beyond six params, stream-vs-loop, complexity, AOP-for-cross-cutting-only) apply verbatim and are not repeated here. The J-standards below are the language doctrine on top of them: each is one rule sentence plus its rationale, enforced at review. A violation is a review finding, not a style comment.

Type-driven design

J1 — Value types are records. Every value type is a record; its compact constructor validates and defensively copies (List.copyOf) so no invalid or externally mutable instance can exist. A final-fields-plus-getters class that could be a record is a review flag: it hand-writes the equality and accessors the language now provides, and hand-written versions drift.

// BAD — hand-rolled value class: the caller's list mutates the "immutable" order later
public final class Order {
    private final List<LineItem> items;
    public Order(List<LineItem> items) { this.items = items; }
    public List<LineItem> items() { return items; }
}

// GOOD — record; the compact constructor validates and copies once, at the boundary
public record Order(OrderId id, List<LineItem> items) {
    public Order {
        Objects.requireNonNull(id);
        if (items.isEmpty()) throw new IllegalArgumentException("order needs items");
        items = List.copyOf(items);
    }
}

J2 — Closed hierarchies are sealed and switched without default. Every domain variant set is a sealed interface, and consumption is switch pattern matching with no default branch — so adding a variant breaks compilation at every consumer instead of falling through silently at runtime. A default arm is legitimate only over non-sealed types you do not control.

// BAD — default arm: a new Refunded variant is silently "unknown" in production
String describe(OrderEvent event) {
    return switch (event) {
        case Created c -> "created";
        case Cancelled c -> "cancelled";
        default -> "unknown";
    };
}

// GOOD — sealed set, no default: adding Refunded breaks the build at every consumer
sealed interface OrderEvent permits Created, Cancelled, Refunded {}

String describe(OrderEvent event) {
    return switch (event) {
        case Created c -> "created";
        case Cancelled c -> "cancelled";
        case Refunded r -> "refunded";
    };
}

J3 — Invariants live in types, not comments. A status enum plus nullable fields makes every reader re-derive which fields are legal in which status, and lets no compiler check the answer. Model one sealed variant per state, each carrying exactly the data that is legal in that state.

// BAD — which fields may be null in which status? The answer lives in reviewers' heads
public record Payment(PaymentStatus status,
                      @Nullable Instant capturedAt,
                      @Nullable String refundReference) {}

// GOOD — each state carries exactly its legal data; no illegal combination compiles
public sealed interface Payment permits Pending, Captured, Refunded {}
public record Pending(PaymentId id) implements Payment {}
public record Captured(PaymentId id, Instant capturedAt) implements Payment {}
public record Refunded(PaymentId id, Instant capturedAt, String refundReference)
        implements Payment {}

J4 — Domain identifiers are value records. record UserId(String value) with validation in the compact constructor — never a raw String or long. Raw identifiers let any ID fill any parameter, so an argument swap compiles cleanly and corrupts data; an identifier record turns the same mistake into a type error.

J5 — Public methods return Optional, never null. A null return is an invisible contract every caller must remember; Optional puts absence in the signature where the compiler and NullAway can see it. Optional is a return type only — never a field (it adds a second null-ish state to store) and never a parameter (it forces ceremony an overload avoids).

Nullness

J6 — Every package is @NullMarked. With JSpecify 1.0, non-null is the default and @Nullable the visible, checkable exception — one package-info.java annotation per production package. Spring Framework 7's own APIs are JSpecify-annotated, so the checker sees through framework calls instead of stopping at them.

J7 — NullAway runs in the build as an ERROR. -Xep:NullAway:ERROR on Error Prone ≥ 2.36, with OnlyNullMarked mode for incremental adoption; the full wiring lives in references/build-and-quality-gates.md. This is prevention, not review-time detection: an NPE the compiler could have refused should never reach a reviewer, let alone production.

J8 — Boundary reads are nullable until proven. Map.get, deserialized DTO fields, and JDBC/ORM reads return "maybe absent" whatever the annotations claim; handle absence at the read site — getOrDefault, computeIfAbsent, orElseThrow — not three stack frames later where the reason is gone.

// BAD — NPE at a distance when the tenant is unknown
var region = regionByTenant.get(tenantId);
applyPolicy(region.policy());

// GOOD — absence decided at the read: default it, or fail naming the cause
var region = regionByTenant.getOrDefault(tenantId, Region.DEFAULT);
var strict = Optional.ofNullable(regionByTenant.get(tenantId))
        .orElseThrow(() -> new UnknownTenantException(tenantId));

Immutability

J9 — Immutable by default. Final fields, List.copyOf on every exposed collection, defensive copies of mutable inputs at construction, and no setters on domain types — evolution is a with-style copy or a builder. Immutable objects are safe to share across virtual threads and to cache without inventing a locking design.

J10 — No static mutable state. The only exception is a deliberately designed cache with documented eviction and thread-safety. Anything else is a hidden global: it couples tests to execution order, leaks state across requests, and is invisible at every call site that depends on it.

Concurrency

J11 — Virtual threads are the default for I/O-bound concurrency. One task, one virtual thread, blocking code that reads top to bottom. Reactive stacks are justified only when streaming or backpressure is itself the requirement — not as a general performance posture. Never pool virtual threads: they cost almost nothing to create, and pooling reintroduces exactly the starvation they exist to remove.

J12 — Pinning advice is runtime-versioned. Since JDK 24 (JEP 491), synchronized no longer pins a virtual thread to its carrier; the remaining pinning sources are native frames and class initializers. Diagnose with the JFR jdk.VirtualThreadPinned event before touching any code. The old "replace synchronized with ReentrantLock" advice applies only to pre-24 runtimes — say which runtime you are on before repeating it.

// BAD on JDK 24+ — mechanical rewrite of working code, citing an obsolete rule
private final ReentrantLock lock = new ReentrantLock();
void refresh() {
    lock.lock();
    try { reload(); } finally { lock.unlock(); }
}

// GOOD on JDK 24+ — synchronized does not pin; verify with JFR, not folklore
synchronized void refresh() {
    reload();
}

J13 — ScopedValue over ThreadLocal for context propagation. ScopedValue is final since JDK 25 (JEP 506): immutable, bounded to a scope, inherited by structured subtasks, and impossible to forget to remove(). A new ThreadLocal requires written justification in the PR.

J14 — Structured concurrency for fan-out — as a preview-gated choice. StructuredTaskScope with a Joiner expresses fork/join with correct cancellation, but it is preview (JEP 505 in 25, renamed in 26's JEP 525, expected final in 27). Policy: preview features are OFF in production unless the project explicitly opts in and records the decision. The no-preview fallback: submit each subtask individually and consume completions via an ExecutorCompletionService (or a list of futures polled as they finish) under an explicit deadline; on the first failure, cancel the remaining futures yourself — note that invokeAll cannot do this, since it blocks until every task finishes and surfaces failures only at Future.get. Either way, every fork is joined — a fire-and-forget submit is a defect.

J15 — Every blocking external call has an explicit timeout. HTTP, JDBC, queue, cache — connect and read timeouts stated at the call site or the client builder, never inherited from a library default, because the library default is usually "infinite".

J16 — Modern replacements (rows, not standards):

Legacy Replace with
JNI FFM — java.lang.foreign (final since 22)
Collect-then-restream pipelines Gatherers — windowFixed, scan, mapConcurrent (final since 24)
java.util.Date / Calendar java.time
Reproducible RNG via ThreadLocalRandom SplittableRandom(seed) — ThreadLocalRandom cannot be seeded
finalize() Cleaner
Pattern.compile per call Cached static final Pattern
Service logic static on a DI bean Instance method — substitutable and injectable

Primitive selection, when shared state or coordination is genuinely needed:

Concept Tool When
Thread-safe collections ConcurrentHashMap, CopyOnWriteArrayList Shared mutable state
Atomic operations AtomicReference, VarHandle, LongAdder Lock-free updates
Locks ReentrantLock, StampedLock, ReadWriteLock Fine-grained locking
Synchronizers CountDownLatch, Semaphore, Phaser, CyclicBarrier Thread coordination
Executors newVirtualThreadPerTaskExecutor(), ForkJoinPool Task scheduling
CompletableFuture thenApply, thenCompose, allOf, anyOf Async composition

Build hygiene

J17 — Versions live in exactly one place. Gradle: the version catalog gradle/libs.versions.toml is the ONLY location a version literal may appear. Maven: BOM imports plus <properties>. A version literal inside a dependency declaration is a review flag — two locations for one version is how a build ends up running two versions.

J18 — The quality gate compiles in CI. Error Prone + NullAway as errors and -Xlint:all, with suppressions only as named annotations at the narrowest scope, each carrying its justification. A warning nobody fails on is documentation of a defect, not a gate.

J19 — Dependency integrity is enforced, not assumed. Gradle dependency verification metadata or a Maven lockfile; reproducible builds; OSV-Scanner in CI (preferred over owasp-dependency-check for its false-positive rate — both are acceptable); a CVSS ≥ 7 finding blocks merge; SBOM published via CycloneDX. Wiring in references/build-and-quality-gates.md.

Testing

J20 — JUnit 6. One version across Platform, Jupiter, and Vintage (Java 17 baseline). Naming keeps the house convention: should_<expected>_when_<condition>, or a @DisplayName sentence about observable behaviour. The unit-test toolkit and shape: Mockito 5 (@Mock, @InjectMocks, BDDMockito; mock collaborators, never the type under test), AssertJ fluent assertions (assertThat().extracting(), assertThatThrownBy()), Given-When-Then structure with one assertion concept per test, no interdependence between tests (any order, any subset), and test-data builders over copy-pasted fixture literals. The pyramid holds: many unit tests, fewer integration/slice tests, few end-to-end — inverted pyramids are slow and flaky where they are supposed to be cheap and exact.

J21 — Parameterized tests for input classes; properties for invariants. Input partitions get @ParameterizedTest; algebraic invariants — round-trips, idempotency, ordering, merge-commutativity — get jqwik properties, and each property names the invariant it defends (serialization_round_trips_any_valid_order), because a property that cannot be named is usually not an invariant.

J22 — Architecture rules are tests. ArchUnit encodes the layered dependency direction, no package cycles, domain-packages-import-no-framework, no field injection, no java.util.Date, and a deprecated-symbol freeze. An approved architecture boundary decision lands as an ArchUnit rule in the same PR — the extension-point discipline of ../../../architecture/solution-architect/references/design-for-predictability.md (D6) made executable. Starter class in references/build-and-quality-gates.md.

J23 — Mutation testing gates new code. PIT 1.19.4+ with the pitest-junit5-plugin (JUnit Platform engines are not built into pitest core) and incremental analysis on PR CI: mutation score ≥ 75% on new and changed classes, plus a weekly full run to keep the history honest. Advisory for the first sprint, then blocking. Never chase 100% — equivalent mutants make it unreachable, and the chase produces assertion noise.

J24 — Integration tests run the real thing. Testcontainers-first, with @ServiceConnection under Spring Boot. A mocked repository in an integration test verifies the mock, not the query, and stays green while the SQL is wrong.

Security — boundary classes (each names its guard)

J25 — Bean Validation on every boundary DTO. Every request object carries constraint annotations and is validated (@Valid) at the controller; an unvalidated request object reaching a service is BLOCKING at review. The guard is the constraint set plus the test that rejects an invalid payload.

J26 — Native serialization of untrusted data is prohibited. ObjectInputStream over anything an outsider can influence is remote code execution waiting for a gadget chain. Where legacy makes it unavoidable, the guard is a JVM-wide ObjectInputFilter allowlist (JEP 290/415) — named in the ticket, asserted by a test.

J27 — Path traversal: resolve, normalize, prove containment. Every path built from external input goes through resolve(...).normalize() and a startsWith(baseDir) check — including the zip-slip variant, where the hostile name arrives as an archive entry.

// BAD — "../../etc/cron.d/job" in an entry name writes outside baseDir (zip-slip)
Path target = baseDir.resolve(entry.getName());
Files.copy(zip.getInputStream(entry), target);

// GOOD — normalize, then prove the result is still inside the base
Path target = baseDir.resolve(entry.getName()).normalize();
if (!target.startsWith(baseDir)) {
    throw new UnsafeArchiveEntryException(entry.getName());
}
Files.copy(zip.getInputStream(entry), target);

J28 — SSRF: allowlist before fetch. A client-supplied URL passes a host allowlist and a deny of loopback, private and link-local ranges (127.0.0.0/8, 10/8, 172.16/12, 192.168/16, 169.254/16, ::1, fd00::/8, fe80::/10) before any request is made — after DNS resolution, so a hostname pointing into the private range is caught too. The guard is the checked resolver the HTTP client is built with.

J29 — ReDoS: regexes over user input are bounded. No nested quantifiers or overlapping alternation over user input; input length bounded before matching; patterns cached as static final. A backtracking regex plus one pathological string is a CPU denial of service.

J30 — Secrets never touch code or logs. Environment or secret manager only; never in source, never logged. Configuration classes' toString redacts secret fields, and a test asserts the redaction — the test is the guard that survives refactors.

Deep-dive references (load on demand)

  • references/build-and-quality-gates.md — the complete build wiring: version catalog, Error Prone + NullAway for Gradle and Maven, the @NullMarked package template, the ArchUnit starter class, PIT configuration, Testcontainers templates, verification metadata, OSV-Scanner, CycloneDX. Load when setting up or reviewing build/CI for any Java module.

Cross-links (owned elsewhere — do not duplicate here):

  • Spring, persistence, API design, messaging — /be's ../../backend/java/backend-developer/references/java-expertise.md; the role skill itself is ../../backend/java/backend-developer/SKILL.md.
  • SQL written anywhere in the stack — the sql language skill, ../sql/SKILL.md.
  • Rust components in the same product — the rust language skill, ../rust/SKILL.md.

Workflow note

This skill owns no gates. The workflow-engine contract and the role agents' gate checks apply unchanged; this skill supplies the language knowledge inside them. Java-specific gate triggers to know (they mirror workflow.yaml — that file decides):

  • New dependencies, new modules, and new service boundaries are ARCH triggers (new_dependency, new_service, cross_boundary).
  • Security-sensitive surfaces — auth, secrets, deserialization, file or URL handling fed by external input — are SECOPS triggers.

Whether those gates fire is the workflow-engine's decision, not this skill's. What this skill supplies is the evidence the gates consume: the sealed model, the stated null contract, the ArchUnit rules, and the green quality gate are what /rev and /verify check against.

Checklist

Before Implementing

  • Domain model sketched as records and sealed types — illegal states unrepresentable (J1–J3)
  • Null contract stated: which packages are @NullMarked, which fields/returns are @Nullable (J6–J8)
  • Test list written — behaviour sentences before any implementation (J20–J21)
  • Stack versions confirmed against ## Versions — re-verified if that table is stale

Before Commit

  • Error Prone + NullAway clean at ERROR (J7, J18)
  • ArchUnit suite green (J22)
  • All tests green, including Testcontainers integration tests (J24)
  • No version literal outside the catalog / properties (J17)
  • Change verified as landed — behaviour observed, not assumed (the verify-landed process skill)

Anti-Patterns to Avoid

Anti-Pattern Problem Solution
Boolean-plus-nullable state Illegal combinations representable; every reader re-derives the legal ones One sealed variant per state (J2, J3)
default arm on a sealed switch New variants are swallowed silently at runtime Exhaustive switch, no default (J2)
@Autowired field injection Hidden dependencies; untestable without a container Constructor injection; ArchUnit forbids the field form (J22)
Raw String/long domain IDs Any ID fits any parameter; swaps compile and corrupt data Identifier records validating on construction (J4)
null return from a public method Invisible contract; NPE at a distance Optional return (J5)
Optional as field or parameter A second null-ish state in storage; ceremony at call sites Optional as a return type only (J5)
ThreadLocal for request context Leaks across carriers; error-prone cleanup ScopedValue (J13)
Pooled virtual threads Reintroduces the starvation virtual threads remove One task, one virtual thread (J11)
Pre-24 pinning advice applied to 24+ Churn with no effect; hides the real pinning sources JEP 491 applies; diagnose with JFR jdk.VirtualThreadPinned (J12)
Version literals in build files Two locations for one version become two versions Catalog / BOM + properties only (J17)
Mocked repositories in integration tests Verifies the mock while the SQL is wrong Testcontainers (J24)
Unvalidated boundary DTO Malformed input reaches domain logic Bean Validation + @Valid, with a rejection test (J25)
Native serialization of untrusted input Gadget-chain remote code execution Prohibited; legacy behind an ObjectInputFilter allowlist (J26)
Catching NullPointerException as flow control Hides the defect the checker would have refused Fix the null contract (J6–J8)