GRILL
Goal
Prove a target is not relying on unsupported claims, fake coverage, or unsafe failure semantics. This is read-only unless repair is explicitly requested.
Workflow
Inspect the target, implementation, tests, history, CI identity, runtime evidence, and project invariants. Attack races, partial failure, restart, malformed input, authorization, retries, rollback, observability, coupling, migration, performance, and speculative abstractions. Questions are not findings; inspect evidence.
For each important problem report:
claim | challenge | evidence | severity | counterexample | required proof/test | correction
When an independent critic uses a selected model/version/variant, invoke it only
through opencode-worker. Bound critic count and reconcile disagreements against
repository evidence.
End with SURVIVED, FAILED, or CONDITIONALLY SURVIVED, followed by concrete
reasons and unmet proof obligations.