audit-third-party-contracts

Verify every third-party and sharp-edged JDK API usage against the contract the upstream documentation actually states

SKILL.md

GitHub

Audit uses of external libraries and sharp-edged JDK APIs against the contract the upstream documentation actually states — not the contract the call site assumes. This lens is distinct from the cache-correctness lenses: the bugs are in the seam between our code and someone else's semantics (dispose-on-error, merge-less collectors, empty timed buffers, live-view iteration windows).

Scope: examples/*/src/main/java/, simulator/src/main/java/, jcache/src/main/java/, guava/src/main/java/. The core is nearly dependency-free; include it only for JDK sharp edges.

Step 1: Enumerate touchpoints

From gradle/libs.versions.toml and each module's build file, list the third-party libraries in scope. Grep each module for their package imports, plus these JDK sharp edges: Collectors.toMap/groupingBy, CompletableFuture composition and cancellation, live collection views held across a time window, ServiceLoader/TCCL-dependent lookups, Object.wait/notify handshakes.

Step 2: Extract the assumed contract

For each touchpoint, state what the call site implicitly assumes about:

  • Error path: what happens to the pipeline/subscription/future when user code or the library throws? Is the failure surfaced or swallowed?
  • Degenerate input: duplicate keys, empty batch, null element, zero duration, single element
  • Cancellation/disposal: is it terminal? who observes it? can the structure be reused after?
  • Threading: which thread runs the callback; what ThreadLocal, ordering, or reentrancy assumptions ride on that?
  • Resource lifecycle: close/dispose idempotency, unsubscribe/leak behavior

Step 3: Verify against the upstream contract

Quote the authoritative source — library javadoc for the version pinned in libs.versions.toml, the reactive-streams specification rule, or JDK javadoc. WebFetch the docs when unsure; do not rely on recall. A finding must cite the upstream sentence that contradicts the call site.

High-yield contract families:

  • Reactive streams (RxJava/Reactor): onError is terminal and disposes; a subscriber without an error handler routes to a global hook and silently unsubscribes; timed buffers emit empty lists on every period; a cancelled sink rejects all future emissions; a subject with no observers drops emissions silently
  • Collectors.toMap throws IllegalStateException on duplicate keys without a merge function
  • CompletableFuture: exceptions in *Async stages surface only via the returned future; cancel() does not interrupt the running task; a throw inside whenComplete replaces/suppresses in non-obvious ways
  • Live vs snapshot views: a live keySet/values passed to another operation observes concurrent mutations; removeAll(null) throws NPE
  • ORM/config libraries (Hibernate, Typesafe Config): session/entity lifecycle and reload semantics the integration assumes. In HOCON an unquoted dotted key is a path (nested objects), and Java-properties keys (-D, .properties) split on every . with no quoting, so how the integration addresses a key (literal quoted key or path) decides which user configurations resolve. When that addressing changed, run the example and user-style configurations against the previous release's jar and the current one. Price it on the names the framework actually looks up: Hibernate names a region by the entity's class name by default, and its region_prefix and legacy timestamps-region names are dotted even when the user's are not

Step 4: Construct the violating scenario

For each contradiction, produce the concrete input or event sequence and the observable damage (hung future, silent data loss, permanent pipeline death, corrupted secondary structure, useless periodic load). Findings without a constructible scenario are suspicions — label them per the auditor contract.

Classify findings with the external-contract category from .claude/docs/finding-taxonomy.md. Also check the module's tests: if the failing scenario is untestable because tests only exercise happy paths (distinct keys, no failures, non-empty batches), report the test gap alongside the bug.