12 - Testing Strategy¶

This document specifies the test plan, test cases, mocking approach, and testing conventions.

Test Framework and Dependencies¶

Dependency	Purpose
JUnit 5 (Jupiter)	Test framework
AssertJ	Fluent assertions
Mockito	Mocking `ChatLanguageModel`, tools, etc.
SLF4J Simple or Logback Test	Logging during tests

Test Conventions¶

Test class naming: {ClassName}Test.java for unit tests, {Feature}IntegrationTest.java for integration tests
Test method naming: testMethodName_scenario_expectedBehavior or descriptive @DisplayName
Structure: Arrange / Act / Assert pattern
No test order dependencies: Each test is independent and self-contained
No real network calls: All LLM interactions are mocked
Deterministic: No reliance on time, randomness, or external state

Test Directory Structure¶

agentensemble-core/src/test/java/io/agentensemble/
  AgentTest.java                              # Agent builder validation
  TaskTest.java                               # Task builder validation
  EnsembleTest.java                           # Ensemble validation (run() preconditions)
  agent/
    AgentExecutorTest.java                    # Agent execution with mocked LLM
    AgentPromptBuilderTest.java               # Prompt construction
  workflow/
    SequentialWorkflowExecutorTest.java       # Sequential execution logic
  tool/
    LangChain4jToolAdapterTest.java           # Tool adaptation
    ToolResultTest.java                       # ToolResult factory methods
  config/
    TemplateResolverTest.java                 # Template variable substitution
  integration/
    SequentialEnsembleIntegrationTest.java    # End-to-end sequential ensemble
    ToolUseIntegrationTest.java               # Agent with tools, mocked LLM
    ErrorHandlingIntegrationTest.java         # Error scenarios end-to-end

Mocking Strategy¶

Mocking the LLM¶

LangChain4j's ChatLanguageModel is mocked using Mockito:

ChatLanguageModel mockLlm = mock(ChatLanguageModel.class);

// For simple text response (no tools):
when(mockLlm.generate(anyList()))
    .thenReturn(Response.from(AiMessage.from("Mocked response text")));

// For tool call response:
when(mockLlm.generate(anyList(), anyList()))
    .thenReturn(Response.from(AiMessage.from(
        ToolExecutionRequest.builder()
            .name("web_search")
            .arguments("{\"input\": \"AI trends 2026\"}")
            .build())))
    .thenReturn(Response.from(AiMessage.from("Final answer after tool use")));

Mocking Tools¶

For AgentTool implementations:

AgentTool mockTool = mock(AgentTool.class);
when(mockTool.name()).thenReturn("web_search");
when(mockTool.description()).thenReturn("Search the web");
when(mockTool.execute(anyString())).thenReturn(ToolResult.success("Search results..."));

Unit Test Cases¶

AgentTest.java¶

Test	Description
`testBuild_withAllFields_succeeds`	Build with every field set, verify all values
`testBuild_withMinimalFields_succeeds`	Build with only role + goal + llm
`testBuild_withNullRole_throwsValidation`	Null role throws `ValidationException`
`testBuild_withBlankRole_throwsValidation`	Blank/empty role throws `ValidationException`
`testBuild_withNullGoal_throwsValidation`	Null goal throws `ValidationException`
`testBuild_withBlankGoal_throwsValidation`	Blank goal throws `ValidationException`
`testBuild_withNullLlm_throwsValidation`	Null LLM throws `ValidationException`
`testBuild_withZeroMaxIterations_throwsValidation`	maxIterations=0 throws
`testBuild_withNegativeMaxIterations_throwsValidation`	maxIterations=-1 throws
`testBuild_withInvalidToolObject_throwsValidation`	Object that is not AgentTool and has no @Tool methods
`testBuild_withAgentToolAndAnnotatedTool_succeeds`	Mixed tool types in same list
`testDefaultValues_areCorrect`	Verify verbose=false, maxIterations=25, etc.
`testToBuilder_createsModifiedCopy`	Modify one field via toBuilder, others unchanged
`testToolsList_isImmutable`	Verify tools list cannot be modified after build

TaskTest.java¶

Test	Description
`testBuild_withAllFields_succeeds`	Build with every field set
`testBuild_withMinimalFields_succeeds`	Build with only description + expectedOutput + agent
`testBuild_withNullDescription_throwsValidation`	Null description throws
`testBuild_withBlankDescription_throwsValidation`	Blank description throws
`testBuild_withNullExpectedOutput_throwsValidation`	Null expectedOutput throws
`testBuild_withBlankExpectedOutput_throwsValidation`	Blank expectedOutput throws
`testBuild_withNullAgent_throwsValidation`	Null agent throws
`testBuild_withSelfReference_throwsValidation`	Task in its own context list
`testBuild_withEmptyContext_succeeds`	Default empty context list
`testContextList_isImmutable`	Verify context list cannot be modified after build

EnsembleTest.java¶

Tests call run() which triggers validation before execution.

Test	Description
`testRun_withEmptyTasks_throwsValidation`	No tasks throws
`testRun_withEmptyAgents_throwsValidation`	No agents throws
`testRun_withUnregisteredAgent_throwsValidation`	Task references agent not in ensemble
`testRun_withCircularContext_throwsValidation`	Tasks A->B->A circular reference
`testRun_withContextOrderViolation_throwsValidation`	Task references context task appearing later
`testRun_withUnusedAgent_logsWarning`	Agent in list but no task uses it
`testRun_withValidConfig_executesSuccessfully`	Happy path with mocked LLM
`testRun_withInputs_resolvedTemplates`	Template variables are substituted
`testRun_calledMultipleTimes_independent`	Each run is independent

TemplateResolverTest.java¶

Test	Description
`testResolve_simpleVariable`	`{topic}` -> value
`testResolve_multipleVariables`	`{a} and {b}` -> values
`testResolve_missingVariable_throwsAll`	Reports ALL missing, not just first
`testResolve_escapedBraces_returnsLiteral`	`{{var}}` -> `{var}`
`testResolve_emptyValue_replacesWithEmpty`	`{x}` with value="" -> ""
`testResolve_nullValue_replacesWithEmpty`	`{x}` with value=null -> ""
`testResolve_noVariables_returnsUnchanged`	No placeholders, returns as-is
`testResolve_nullTemplate_returnsNull`	null input -> null output
`testResolve_blankTemplate_returnsBlank`	"" input -> "" output
`testResolve_nullInputs_treatedAsEmpty`	null map treated as empty
`testResolve_sameVariableMultipleTimes`	`{a}{a}` -> "XX"
`testResolve_underscoreInName`	`{under_score}` is valid
`testResolve_extraInputsIgnored`	Unused keys in map are ignored

AgentPromptBuilderTest.java¶

Test	Description
`testBuildSystemPrompt_withAllFields`	Role + background + goal + responseFormat
`testBuildSystemPrompt_withoutBackground`	Background=null, omitted
`testBuildSystemPrompt_withEmptyBackground`	Background="", omitted
`testBuildSystemPrompt_withoutResponseFormat`	responseFormat="", omitted
`testBuildSystemPrompt_minimalAgent`	Only role + goal
`testBuildUserPrompt_withoutContext`	No context, just task + expected output
`testBuildUserPrompt_withSingleContext`	One prior task output
`testBuildUserPrompt_withMultipleContexts`	Multiple prior outputs, order preserved
`testBuildUserPrompt_contextOrderPreserved`	Outputs render in context list order
`testBuildUserPrompt_emptyContextOutput`	Context with empty raw text, headers still shown

LangChain4jToolAdapterTest.java¶

Test	Description
`testAdapt_createsCorrectSpec`	ToolSpecification name, description, parameters match
`testAdapt_execution_delegatesToAgentTool`	Calls execute(), returns output
`testAdapt_executionError_returnsErrorString`	ToolResult.failure -> "Error: ..."
`testAdapt_executionThrows_returnsErrorString`	execute() throws -> caught, returns error
`testAdapt_nullResult_returnsEmpty`	execute() returns null -> ""

ToolResultTest.java¶

Test	Description
`testSuccess_withOutput`	success=true, output set, errorMessage=null
`testSuccess_withNullOutput`	Null coerced to empty string
`testFailure_withMessage`	success=false, errorMessage set, output=""

Integration Test Cases¶

SequentialEnsembleIntegrationTest.java¶

Test	Description
`testTwoTaskEnsemble_contextPassedCorrectly`	Mock LLM returns "Research X", verify second task's prompt contains "Research X"
`testSingleTaskEnsemble_executesSuccessfully`	One task, one agent, verify output
`testThreeTaskChain_allContextFlows`	Task 3 depends on Task 1 and 2, verify both in context
`testTemplateSubstitution_variablesResolvedInPrompts`	{topic} resolved before execution, verify LLM receives resolved text
`testEnsembleOutput_containsAllTaskOutputs`	Verify EnsembleOutput has correct taskOutputs list
`testEnsembleOutput_rawIsLastTaskOutput`	raw field equals the final task's output
`testEnsembleOutput_durationsArePositive`	Duration fields are > 0

ToolUseIntegrationTest.java¶

Test	Description
`testAgentWithTool_toolCalledAndResultUsed`	LLM calls tool, tool returns result, LLM uses result in final answer
`testAgentWithMultipleToolCalls_allExecuted`	LLM calls tool A then tool B then produces answer
`testAgentWithToolError_errorFedBackToLlm`	Tool throws, error message sent to LLM, LLM produces answer
`testAgentWithToolFailureResult_errorFedBack`	Tool returns ToolResult.failure(), error sent to LLM
`testMaxIterations_sendsStopMessages`	LLM always calls tools, verify stop messages sent
`testMaxIterations_throwsAfterThreeStopMessages`	After 3 stops, MaxIterationsExceededException thrown
`testNoTools_directLlmCall`	Agent with no tools, verify single generate() call
`testMixedTools_bothTypesWork`	Agent with AgentTool + @Tool annotated, both called

ErrorHandlingIntegrationTest.java¶

Test	Description
`testLlmThrows_wrappedInTaskExecutionException`	LLM throws RuntimeException, wrapped properly
`testSecondTaskFails_firstOutputPreserved`	Two tasks, second fails, first output in exception
`testMissingTemplateVariable_throwsBeforeExecution`	PromptTemplateException, LLM never called
`testValidationFailure_noTasksExecuted`	ValidationException, no LLM interaction
`testTaskExecutionException_hasCorrectContext`	Exception fields (taskDescription, agentRole, completedOutputs) are correct

Test Execution¶

Tests are run via Gradle:

# Run all tests
./gradlew test

# Run specific test class
./gradlew test --tests "net.agentensemble.AgentTest"

# Run specific test method
./gradlew test --tests "net.agentensemble.AgentTest.testBuild_withNullRole_throwsValidation"

# Run integration tests only
./gradlew test --tests "net.agentensemble.integration.*"

Coverage Expectations¶

All public methods of domain objects: 100% branch coverage
All validation paths: tested with both valid and invalid inputs
All exception types: at least one test that triggers each
Template resolver: all edge cases from the matrix in 07-template-resolver.md
Prompt builder: all conditional sections (with/without background, context, etc.)
Agent executor: tool loop, no-tool path, error handling, max iterations
Workflow executor: context passing, error propagation, output assembly