---

name: pharo-tdd description: Drive a Genie change test-first through the genie MCP tools, working entirely from the live image (no files). Use when adding behavior or fixing a bug in the image — anything where a test should lead.

Pharo TDD loop (live image, via the genie MCP tools)

All work goes through the genie MCP tools — there are no files. Prefer the most specific MCP tool available: read with the read tools, change methods with define_method, change classes with define_class, and use dedicated refactoring/removal tools when they exist. Use eval only as a last resort when no specific MCP tool can express the operation, and keep it minimal. Confirm every change by reading it back. Never assume a class/selector exists — look it up first.

Loop

1. Locate. Find the class and its test class: list_classes (package) → list_methods → get_method_source. Find the matching ...Test (TestCase subclass). If none exists, create it via define_class (superclass TestCase).

2. Red. Write or adjust ONE failing test with define_method (protocol tests). Run it: run_test with class_name. Expect failure. If it passes immediately, the test isn't exercising the new behavior — fix the test before writing code.

3. Green. Implement the smallest method that could pass, via define_method. Run run_test again → expect green. Read the method back with get_method_source to confirm what's installed.

4. Regression check. Run run_test for the whole package_name. A green class test with a red package is not done.

5. Refactor. Clean up while green, re-running run_test after each step. (For larger cleanups, switch to the pharo-refactor skill.)

6. Done. Persist with save_image gated on tests — pass test_package so it runs them and saves the image only if green. Summarize what changed (class >> selector list).

Debugging a red test

Genie returns full stack traces: failing block source, receiver, and its variables. Read them. Reproduce with run_test, read/search tools, and the smallest existing MCP call that exercises the behavior. Use eval only if the needed probe cannot be expressed through a dedicated MCP tool; if you must use it, make it minimal and non-mutating unless the task explicitly requires a mutation. Fix the smallest thing the trace implicates, then re-run the test. Don't guess.

Rules

• One test, one reason to fail, at a time.
• Modern class API only (ShiftClassInstaller / define_class), never the deprecated subclass:instanceVariableNames:classVariableNames:package:.
• Tests are methods too — install them with define_method, protocol tests.
• eval is always a last resort, never the default path.