yogthos/clojure-lora-training/arch-refactor-rounds

Architectural Refactoring in Rounds

When cleaning up a codebase, work in small rounds of 3-5 concrete tasks. Each round is independently verifiable — imports work, tests pass. This keeps the codebase functional at all times and prevents large, unverifiable diffs.

Round Structure

Each round:

1. 3-5 concrete tasks — specific files, specific changes, no ambiguity
2. One task at a time — make the change, verify imports, move to next
3. End-of-round verification — import all changed modules + run relevant tests

Task Categories (in priority order)

Priority order for early rounds:

1. Create missing shared modules (models, types, constants that multiple modules import but that don't exist)
2. Remove noise ([LEGACY] markers, empty files, dead docstrings)
3. Centralize scattered concerns (env vars, config loading, logging setup)
4. Fix import chains (ensure __init__.py files export what's actually used)

Later rounds:

1. Consolidate duplication (merge similar classes, extract shared base)
2. Fix design patterns (god objects, feature envy, tight coupling)
3. Reorganize module boundaries (move files between packages)

Verification Pattern

After each task:

uv run python -c "from changed.module import key_symbol; print('OK')"

After each round:

uv run python -c "
import module1; print('OK')
import module2; print('OK')
...
"

Then run relevant tests. Add -p no:cacheprovider to every pytest invocation during a refactor — it disables the on-disk cache so a stale .pytest_cache can never mask an import failure or reuse a stale collected test:

uv run python -m pytest tests/unit/test_changed_modules*.py -x -q --tb=short -p no:cacheprovider

Full-suite-per-round (when the suite is fast): on this project the full tests/unit suite (~837 tests, 5 skipped) runs quickly, so prefer running the FULL suite every round rather than just the changed modules — stronger regression coverage, and it's the same command as the session-end gate:

uv run python -m pytest tests/unit -q --tb=short -p no:cacheprovider

Session-end gate (non-negotiable): before declaring the refactor done, run the FULL unit suite and require zero failures. If spaCy (en_core_web_sm) is NOT installed, add --ignore=tests/unit/test_content_classifier.py --ignore=tests/unit/test_transfer.py to drop the spaCy-dependent tests (see Pitfalls); if spaCy IS installed, run everything.

Pitfalls

• git push origin main may require human approval via the approval_provider. If blocked, tell the user explicitly and do not consider the session complete.
• Tests in test_content_classifier.py and test_transfer.py require spaCy (en_core_web_sm). When spaCy isn't installed, skip those with --ignore=tests/unit/test_content_classifier.py --ignore=tests/unit/test_transfer.py.

What NOT to do in early rounds

• Don't refactor logic — only move/rename/delete dead code
• Don't "clean up" code you're not targeting — stay focused on the task list
• Don't skip verification — one broken import cascades
• Don't do more than 5 tasks per round — verify and commit before continuing