Exercise picker with prompt, hints, and fixed input IR
Starter code editor
One-click validation
Optional "show solution" button
Subprocess-based evaluation (safer for runtime crashes / SystemExit)
Supports solve(input_ir), main([input_ir]), or top-level script execution
Validators can inspect return values and/or stdout/stderr
Metadata-driven checks in TOML (contains/regex/exception/entrypoint)
Declarative sidecar files per exercise:
- <prefix>.input.ll (optional; omit for from-scratch exercises)
- <prefix>.starter.py
- <prefix>.solution.py
- <prefix>.validator.py (optional)

Files

exercises/poc/exercises/*.toml: exercise metadata (references files + checks)
exercises/poc/exercise_bank.py: TOML loader
exercises/poc/evaluator.py: execution + metadata checks + custom validators
exercises/poc/portal.py: local web server and UI
exercises/poc/exercises/*.validator.py: define validate(run, exercise, helpers)

Notes

This is intentionally minimal, not production hardened.
Validation is deterministic and workshop-focused.
For IR tasks, validators accept either returned IR text or printed IR text when allow_stdout_ir is enabled for the exercise.
starter_code and solution_code are loaded from: <prefix>.starter.py and <prefix>.solution.py.
input_ir is loaded from <prefix>.input.ll when present; if omitted, the exercise has no input module and users build from scratch.
Validator files are first-class: keep exercise-specific logic in <prefix>.validator.py instead of routing through a global name map.
id is inferred from filename stem (do not set it in TOML):
- 01_F01.toml -> F01
- 10_B02.toml -> B02
Custom validators are discovered by sidecar name only:
- <prefix>.validator.py