This is a useful distinction between functional and structural correctness. In a long coding-agent run, I’d make the constraint set explicit and layered: immutable API contract, a short architectural checklist, then repo-local examples. Keeping a compact state summary between tool calls seems safer than replaying every log line; otherwise the important invariant gets diluted by test output.
I’d be curious whether the failures were mostly from lost constraints in the prompt or from the agent seeing contradictory evidence in the repository. A verifier that reports the first violated invariant might improve recovery more than another generation pass.