The Wrapper That Earns Its Keep
Before touching any code I stopped and asked whether the casehub Agent class should exist at all. LangChain4j has its own agent model — UntypedAgent, AgenticScope, supervisor patterns. Was the casehub version just reinventing something the library already provides better?
The answer turned out to matter more than the refactoring.
The casehub Agent does one thing LangChain4j doesn’t: it applies jq transformations to the case context before feeding data to the LLM, and to the LLM’s output after getting it back. An inputSchema: "{ title: .caseTitle, docs: .documents }" extracts just what the model needs from the full case context. An outputSchema: "{ decision: .verdict }" writes only the relevant field back. LangChain4j handles the LLM call; casehub provides the bridge between its context model and whatever the LLM expects.
That’s genuine value. The class stays — but it should be jq-agnostic. Agent was holding two JqTransformer fields, coupling it directly to a specific jq implementation. The real contract is a function: UnaryOperator<JsonNode>.
The change in Agent is one line per field. The change in AgentBuilder is more interesting: inputSchema(String) still works (it creates a JqTransformer internally and wraps ::apply), but now inputTransformer(UnaryOperator<JsonNode>) gives CDI callers a way to pass a JQEvaluator-backed lambda. Defaults to identity when neither is set.
final UnaryOperator<JsonNode> resolvedInput =
inputSchema != null ? new JqTransformer(inputSchema)::apply
: (inputTransformerFn != null ? inputTransformerFn : UnaryOperator.identity());
Writing the test for AgentBuilder hit a wall immediately. ChatModel can’t be stubbed as a lambda — "ChatModel is not a functional interface". Every method on it, including chat(ChatRequest) (the documented primary API), is default. Claude found doChat(ChatRequest) by reading the source: it’s the real extension point, also default, throwing RuntimeException("Not implemented") unless overridden. The chat method orchestrates logging and delegates to doChat. Override doChat, not chat.
The second issue on this branch — CommandContent — was straightforward after the main work. WorkerScheduleEventHandler.dispatchCommand() was building its COMMAND content as a raw HashMap with string keys. Replacing it with a record makes the wire format visible:
record CommandContent(
String type, String capability, String correlationId,
Map<String, Object> input, String deadline) {}
@JsonInclude(NON_NULL) on the record handles the optional deadline without branching. The wire format is unchanged — claudony reads JSON, not the Java type.
The bigger picture from this branch: the jq transformation layer is what makes AI workers composable with the case engine. The agent doesn’t need to know what a case is. The case engine doesn’t need to know how the model formats its output. The transformer between them is the seam.