Constraint kinds
A constraint kind classifies a requirement extracted from your Python code: must_raise, boundary, enum_valid and so on. The kind decides which rule generates the test and how the verifier violates the requirement to prove the test can fail. This page lists every kind defined in the Python package, which spec reader produces it, and whether the rule engine handles it.
What are the kinds?
| Kind | Meaning | Typical source | Rule engine | Violation injected |
|---|---|---|---|---|
must_raise | Raises an exception under a condition | Raises: docstring entry, try/except re-raise, standalone raise | yes | raise replaced with pass |
boundary | Value above, below or at a threshold | if x <= 0: raise, assert x > 0, Field(gt=0), min_length, max_length | yes | threshold literal replaced with -9999 |
enum_valid | Value must be one of a set | if x not in VALID: raise, Literal[...] | yes | raise replaced with pass |
enum_invalid | An invalid value must be rejected | defined in the model; no reader emits it in the tested code | no | none |
must_return | Returns a specific value or type | Returns: docstring | yes | non-None returns become return None |
not_none | Return must not be None | defined in the model; no reader emits it in the tested code | no | none |
not_null | A variable (or PySpark column) must not be null | if x is None: raise | yes | raise replaced with pass |
type_check | isinstance guard | if not isinstance(x, ...): raise | yes | raise replaced with pass |
auth_check | Authentication or permission guard | if not request.user.is_authenticated: raise | no | raise replaced with pass |
bare_except | Bare or broad except that swallows everything (a smell) | except Exception: pass | no | none |
silent_fail | Returns None instead of raising | if not result: return None | yes | first silent return becomes a raise |
magic_value | Hardcoded string or number in a condition | if status == "admin": | yes | raise replaced with pass |
mutation | A mutant that survived | mutmut results (optional) | no | mutmut apply |
bug_repro | Reproduce a reported bug | quell reproduce | yes | none (already broken) |
custom | Free-form requirement | LLM-handled | yes | raise replaced with pass |
The "Rule engine" column says whether can_handle is true for the kind. Requirements the rule engine cannot handle go to the LLM only when you pass --use-llm, otherwise they are scaffolded or flagged with a reason.
What cannot the rule engine do?
It returns no test when all required parameters are complex objects, when the guard checks self.attr (it needs class instantiation) or a local variable such as a database result (it cannot be injected through a stub), and for async functions it wraps the call with asyncio.run. FastAPI and Flask route handlers go to a framework engine that needs an app object found in the project.
Which readers produce which kinds?
| Reader | Kinds |
|---|---|
Guard clauses (quell find) | boundary, not_null, enum_valid, type_check, auth_check, bare_except, silent_fail, magic_value, must_raise |
| Docstrings | must_raise, boundary, enum_valid, must_return |
| Types and Pydantic models | boundary, enum_valid |
| Bug reports | bug_repro |
See how it works and the five-gate pipeline.