In long-running work, one writer changes the code and a read-only reviewer reports findings, and they alternate until a round finds nothing worth fixing.
Leaves out
The loop is for long-running work and not for every change, and its reviewer is another agent.
Fails
When the reviewer's own severity grades are accepted, and a loop closes with wrong answers filed as minor.
Evidence
The grimoire skill, which describes the loop and its stopping rule
Tool: Oracles, mutants and control cases
Coverage
In places
Does
Where a reference exists, a harness checks output against it. The install script of fuz and the fuz_schema library also have their code broken on purpose to show that a test catches it, and the divergence detector of tsv needs a negative case for each pattern, to prove it can report a failure.
Leaves out
Most code has no reference to compare against, and only those two have a mutation run over their tests.
Fails
When both sides agree on the same bug, when the comparison never looks at the property that is wrong, or when a harness has no control case and reports success for reasons unrelated to the code.
Cost
A twin roughly doubles what is written by hand.
Wanted
Routine mutation and branch-coverage runs for tsv.
Evidence
The fuz installer's mutation gate, which breaks one behavior at a time and checks that a test catches it
fuz_schema's mutants, where a survivor needs a written reason
The tsv divergence detector's overmatching audit, which requires a negative case for each pattern
A failing fixture, lint or case is made to pass by editing the expectation, suppressing the lint or skipping the case.
Session: The expected diff is read as code
Coverage
In places
Does
After an update task regenerates the expected files of a fixture test, the session reads their diff as it would read code, since the diff is the assertion.
Leaves out
Only a fixture test has such a diff to read.
Evidence
The testing reference on fixture tests, where the diff of the expected files is reviewed as code
Written: Fix the code, not the expectation
Coverage
In places
Does
The written rules say an expected file is generated and never edited by hand, a Rust lint is suppressed only with a stated reason, and a difference between twins is recorded and never skipped by hand. tsv adds that a correct fixture is never changed to make a test pass.
Leaves out
No rule covers a suppression in TypeScript or a skipped test.
Fails
When the expectation really is wrong, and telling that apart from a convenient edit is left to the session.
Evidence
The fuz-stack skill's rules by topic, where expected files are generated and never hand-edited
The Rust lint rules, on suppressing a lint with a reason
Tool: Suppressions that expire
Coverage
In places
Does
A Rust lint suppressed with an expectation warns once the lint stops firing, and the gate treats the warning as an error. A recorded difference between twins fails once the difference is gone. The pinned counts of tsv fail when a count moves, and its fixtures are compared against a live, pinned Prettier, so a fixture edited by hand fails.
Leaves out
Where no outside reference exists, an expected file regenerated from wrong output passes, and so does a suppression with a plausible reason. The record of twin differences is checked only when the cross-backend suites run, which for the Rust side is on demand.
Fails
When the reason on a suppression is accepted without being checked against the code.
Evidence
The Rust lint rules, where a suppression warns once its lint stops firing
The twin reference on enforcement, where a recorded difference fails once it stops showing
A check skips, caps its run or falls back, and still shows green.
Written: Read the exit code
Coverage
Throughout
Does
The written rules make the gate's exit code the pass condition, never its filtered output.
Fails
When a gate's output is piped through a filter that hides its exit code, and the commit follows anyway.
Evidence
The fuz-stack skill on the gate, which says to read the gate's exit code
The Rust gate, whose pass condition is the exit code
Tool: Suites that fail on a skip
Coverage
In places
Does
The suites of zap fail when a program a test needs is missing, unless an explicit allowance names it. fuz fails when a database suite is left off its CI list, pglet's CI fails when a gate reports nothing, and the pinned counts of tsv fail when a count moves.
Leaves out
Elsewhere a skip may only warn, as when the install suite of fuz finds a program it needs missing.
Fails
When a skip stays green with a warning that nobody reads.