BDD Test Generation with Tzatziki

Generate valid Cucumber .feature files from a user's functional specification, using Tzatziki's step definition library as the source of truth for legal step patterns.

Principles

These explain the reasoning behind the workflow — understanding them helps you handle edge cases the workflow doesn't explicitly cover.

1. Steps come from code, not imagination. Tzatziki provides hundreds of pre-built @Given / @When / @Then patterns in its *Steps.java files. Inventing step text that doesn't match a real definition produces UndefinedStepException at runtime. Per-module reference files in references/steps-*.md contain every legal step pattern — read the relevant ones before writing any scenario.

   ⚠️ Never paraphrase or improvise step keywords. Many step patterns use regex constants like COMPARING_WITH that expand to a small fixed set of keywords (e.g., "" empty, " exactly", " at least", " only"). Do not substitute natural-language synonyms such as "containing", "including", "matching", or any other word that "sounds right" — if the word does not appear in the @Given/@When/@Then annotation regex, the step will be undefined at runtime. When in doubt, re-read the Java source in references/steps-*.md rather than guessing.

2. Verify the environment before writing tests. Run at least one existing Cucumber test in the target module before creating new feature files. This catches missing dependencies, broken bootstrap, or misconfigured runners early — before you've invested effort in writing scenarios that can't execute. If no test exists yet, create the minimal bootstrap first and confirm it discovers at least one scenario.

3. YAML by default for structured data. Tzatziki scenarios are most readable when request bodies, database fixtures, Kafka payloads, and expected results use """yml doc strings. Fall back to Gherkin tables for naturally tabular data, or raw JSON only when the contract requires it.

4. Cover exactly what the user asked for — then help them think about what they missed. Generate scenarios for every functional behavior in the user's specification — not just the happy path. But don't silently add extra scenarios either. Instead, after covering the requested scope, actively identify edge cases and present them to the user as optional additions. To identify edge cases effectively, look at three things: (a) every external service call in the scenario — what happens if it returns an error (4xx, 5xx) or times out? (b) every data collection — what happens if it's empty or contains unexpected values? (c) existing test files in the project that test similar features — they often contain error-handling patterns you can adapt for the new scenario. The user decides which edge cases to include, but your job is to surface them so nothing important is missed.

5. Reuse what exists. If the project already has a runner, bootstrap class, feature location convention, or glue configuration — reuse them. Creating duplicates causes classpath conflicts and confuses test discovery.

6. Prefer updating existing scenarios over creating new ones. When a specification adds a new field, header, or assertion to an already-tested behavior (e.g., "add a traceId header to the message this flow already publishes"), the right approach is almost always to modify the Then assertion of an existing scenario that already exercises that behavior, rather than creating a brand-new scenario that duplicates all the same Given/When setup just to check one additional field. Creating a new scenario is only justified when the behavior itself is genuinely new — a different trigger, a different code path, or a different precondition combination that no existing scenario covers. This matters because test suites grow quickly, and every duplicated scenario is a maintenance burden: when the shared setup changes, every copy must be updated in lockstep or the suite becomes inconsistent.

Workflow

1. Understand the Specification

Analyze the user's input and break it into a checklist of distinct functional behaviors. Each behavior will become one or more scenarios. If anything is ambiguous, ask before proceeding.

While analyzing, classify each behavior as one of:

• New behavior — a genuinely new trigger, code path, or precondition combination that no existing scenario covers. This requires a new scenario.
• Additive change — a new field, header, assertion, or output added to an already-tested behavior (e.g., "add a traceId header to the message this scenario already publishes"). This should be handled by modifying the existing scenario's assertions, not by creating a new scenario. See Principle 6.

Also extract:

• External dependencies — every API, service, or data source the feature interacts with. Each one is a potential source of edge-case scenarios (errors, timeouts, empty responses).
• Performance or reliability hints — if the spec mentions performance concerns, large data volumes, or error handling, note them. These signal scenarios the user likely cares about even if they didn't write explicit acceptance tests for them.
• Implicit error paths — if the spec says "call service X to get Y", the spec is describing the happy path. The failure path (service X returns an error) is an edge case to suggest.

2. Discover the Project

Detect the build tool (prefer wrappers: ./mvnw over mvn, ./gradlew over gradle), then:

1. Always read references/steps-core.md — core step definitions (ObjectSteps) are used in every Tzatziki project for variables, assertions, and data manipulation.
2. Detect which tzatziki-* modules the project depends on by inspecting pom.xml or build.gradle for dependency declarations. A quick way to extract them:

   grep -o 'tzatziki-[a-z-]*' pom.xml | sort -u

3. Read the matching per-module reference(s) based on detected dependencies:
   • HTTP/REST testing → references/steps-http.md
   • Spring context → references/steps-spring.md
   • JPA/database → references/steps-spring-jpa.md
   • Kafka messaging → references/steps-spring-kafka.md
   • MongoDB → references/steps-spring-mongodb.md
   • OpenSearch → references/steps-opensearch.md
   • Logging assertions → references/steps-logback.md
   • MCP/AI testing → references/steps-mcp.md
4. When in doubt between a step pattern inferred from a .feature example and one read from the Java *Steps.java source, trust the Java source.
5. Catalog error-handling patterns from existing .feature files you encounter during discovery. Note how the project tests HTTP error responses (e.g., mocking 404/500 from external APIs), empty data, cache misses, or retry behavior. You'll use these patterns as templates when suggesting edge cases in Step 5.

3. Check the Bootstrap

Search for the existing test infrastructure:

• Runner: @Suite + @IncludeEngines("cucumber") (JUnit 5) or AbstractTestNGCucumberTests
• Feature location: @SelectClasspathResource(...) or @SelectDirectories(...)
• Spring context: @CucumberContextConfiguration + @SpringBootTest (if Spring is used)

If any piece is missing, plan to create it. Read references/bootstrap-templates.md for the JUnit 5 runner and Spring configuration templates.

4. Verify the Environment

Run the existing test suite in the target module to confirm it works:

# Maven
./mvnw -pl <module> -Dtest=<RunnerClass> test
# Gradle
./gradlew :<module>:test --tests <RunnerClass>

A passing or functionally-failing run is fine — what matters is that Cucumber discovers and executes scenarios. Tests run: 0 means the runner or feature discovery is broken and needs fixing before you write anything new.

5. Propose the Plan

⛔ This is a mandatory checkpoint. Do not proceed to Step 6 without the user's explicit approval. Even if the user's specification already includes detailed acceptance tests, present the plan anyway — the user expects to review it, may want to adjust scope, and needs the chance to approve or add edge cases they care about. Skipping this step risks generating the wrong tests and wasting effort.

Output the full plan as regular response text first (never embed a long plan inside the ask_user question — the UI truncates it). Include:

• Files to create or modify
• A table mapping each requested functional behavior to a scenario — for each row, indicate whether this is a modify (updating an existing scenario) or a new scenario. Use the behavior classification from Step 1 to decide.
• Any bootstrap work needed
• Suggested edge cases (clearly marked as optional) — see below for how to identify them

Prefer modifying existing scenarios

Before proposing a new scenario, search the project's existing .feature files for scenarios that already exercise the same behavior flow (same endpoint, same Kafka topic, same trigger). When you find one, the default action is to modify that scenario — for example, adding a new field to its Then assertion table or adding a header check — rather than creating a new scenario with duplicated Given/When steps. This avoids test bloat and maintenance overhead.

Create a new scenario only when:

• No existing scenario covers the relevant behavior flow at all
• The new behavior requires genuinely different preconditions (different data setup, different trigger, different error path) that would make the existing scenario confusing if combined
• The user explicitly asks for a separate scenario

Example — adding a traceId header to an existing emitted message:

• ❌ Wrong: Create a new scenario "published message includes traceId header" that copies the entire Given/When from the existing message-publication scenario, just to add headers: traceId: abc-123
• ✅ Right: Modify the existing message-publication scenario's Then block to include headers: traceId: abc-123 alongside its existing assertions

Then use the ask_user tool with a short, focused question asking only for the user's decision (e.g. "Does this plan look good, and which optional edge cases would you like to include?"). Do not write any .feature file until the user confirms.

Identifying edge cases

Use the external dependencies, error patterns, and spec hints you gathered in Steps 1 and 2 to build a concrete list of optional edge-case scenarios. Work through these categories:

1. External service failures — For each external API call in the scenario consider what happens when it returns:

   • A client error (400, 404) — e.g., the resource doesn't exist
   • A server error (500, 502, 503) — e.g., the service is down
   • An empty or unexpected response body

2. Empty and boundary data — For each data collection involved:

   • What if the collection is empty? (e.g., a super model with no models, a model with no articles, an overridden rule with an empty value list)
   • What if it contains a single element vs. many?
   • What if values are null or missing?

3. State and ordering — Cache hit vs. miss, pre-existing vs. fresh data, concurrent modifications, retry after failure (CREATED → TO_RETRY → ERROR).

4. Patterns from existing tests — In Step 2 you cataloged error-handling patterns from the project's existing .feature files. Adapt those patterns to the new feature. For example, if existing tests mock a 404 from /masterdata/v2/arbo/models/{id}/articles and assert "0 kafka messages", suggest the same pattern for any new external API the feature calls.

Present each suggested edge case with a one-line description and mark them as (optional). The user decides which to include.

6. Implement and Validate

⛔ Mandatory: run tests before responding. After writing or modifying any .feature file, you must run the tests and confirm that the output contains zero "step(s) are undefined" / UndefinedStepException errors before you respond to the user. Do not present a summary table, declare "implementation complete", or ask for feedback until this validation has passed. A visual check of the step text is not sufficient — only the test runner output is authoritative.

1. Write the .feature file using exact step patterns from step 2 and matching the existing project's style (scenario naming, Background usage, tags, data format conventions).
2. If no feature convention exists, default to src/test/resources/features with @SelectClasspathResource("features").
3. Run the tests and inspect the output. The test output is the ground truth for step validity — no amount of visual inspection replaces it. Look specifically for these errors:
   • "step(s) are undefined" + "You can implement these steps" — means the step text doesn't match any registered step definition. This is the most common mistake and it must be fixed before the task is considered done. Go back to the per-module references/steps-*.md files, find the correct pattern, and rewrite the step.
   • UndefinedStepException — same root cause, different presentation.
   • ParseError — malformed Gherkin syntax.
   • Tests run: 0 — runner or feature discovery is broken. Accept functional assertion failures (e.g. expected 200 but got 404) — they're expected in BDD when the product behavior isn't implemented yet.
4. Repeat until every requested behavior from step 1 is present and zero undefined-step errors remain in the test output.

Post-implementation modifications

When the user requests changes to already-written scenarios (e.g. convert to Scenario Outline, add an assertion, merge steps, rename, restructure), the validation requirement still applies:

1. Apply the requested edits to the .feature file.
2. Re-run the tests immediately — even if the change looks trivial. Edits that restructure step text, add new Then assertions, or reorganize tables frequently introduce subtle syntax mismatches that only the test runner can catch.
3. If undefined-step errors appear, fix them before responding to the user.

This loop (edit → run → fix) applies to every modification, not just the initial implementation. Never skip the test run, even if the user is asking for a "small" change.

Running a single feature from the CLI: When you need to target one specific feature file or scenario line, use cucumber.features as the selector — not -Dtest=... or cucumber.filter.name. Read references/cli-execution.md for the full details, because cucumber.features triggers standalone Cucumber execution that bypasses the runner's @ConfigurationParameter annotations.

Success Criteria

• At least one existing test was run and confirmed operational before new scenarios were written.
• The plan was presented to the user and approved before any .feature file was written.
• Generated scenarios cover 100% of the functional behaviors the user explicitly requested.
• All step text comes from real Tzatziki step definitions (no invented steps). In particular, no natural-language synonyms were substituted where the step regex expects a specific keyword or an empty string (see Principle 1).
• Tests were run after every .feature edit (initial implementation and subsequent modifications) and the output contains zero "undefined step" errors. This is the definitive validation — if the test output says "step(s) are undefined", the task is not done. The agent must never declare implementation complete or present a summary without first running the tests in the same turn.
• Any required runner or Spring bootstrap files are in place.
• Tests are discovered and executed with the correct build tool command.
• Edge-case scenarios were identified, presented to the user, and included only if approved.
• Any remaining failures are functional assertions, not technical setup problems.

Reference Files

Read these when needed — they contain templates and detailed guidance that would clutter the main workflow:

Per-Module Step References (read based on detected dependencies)

• references/steps-core.md — ObjectSteps: variables, assertions, data manipulation, type conversions. Always read this — core steps are used in every Tzatziki project.
• references/steps-http.md — HttpSteps: HTTP mocking, request/response assertions, REST API testing. Read when the project uses tzatziki-http.
• references/steps-spring.md — SpringSteps: Spring context, properties, bean manipulation. Read when the project uses tzatziki-spring.
• references/steps-spring-jpa.md — SpringJPASteps: JPA entity management, database fixtures, persistence assertions. Read when the project uses tzatziki-spring-jpa.
• references/steps-spring-kafka.md — KafkaSteps: Kafka topic management, message producing/consuming, async assertions. Read when the project uses tzatziki-spring-kafka.
• references/steps-spring-mongodb.md — SpringMongoSteps: MongoDB collection management, document fixtures, query asse/rtions. Read when the project uses tzatziki-spring-mongodb.
• references/steps-opensearch.md — OpenSearchSteps: index management, document indexing, search assertions. Read when the project uses tzatziki-opensearch.
• references/steps-logback.md — LoggerSteps: log assertion, log level management. Read when the project uses tzatziki-logback.
• references/steps-mcp.md — McpSteps: MCP server testing, tool invocation, AI integration. Read when the project uses tzatziki-test-mcp.

Other References

• references/bootstrap-templates.md — JUnit 5 runner and Spring @CucumberContextConfiguration templates. Read when you need to create a new runner or Spring bootstrap class.
• references/cli-execution.md — How to run a single feature or scenario from the CLI using cucumber.features, including the Maven/Gradle property-mirroring mechanism. Read when you need to target a specific .feature file instead of running the full suite.