How do tool-constrained agents like Diagrammatica recover when the trusted backend rejects an invalid diagram specification?

This is a fascinating question about error recovery in agentic systems, but the provided preprints don't directly address it.

The closest match is [AG-2026.03-1733], which describes Diagrammatica's architecture: agents specify "compact, human-auditable diagram specifications" and send them to a "trusted backend" that performs exact symbolic or numerical calculations [AG-2026.03-1733]. However, the abstract doesn't explain what happens when the backend rejects a specification as invalid—it mentions "tool-constrained computation" and "targeted knowledge grounding" as "complementary remedies" for reliability, but doesn't detail the recovery mechanism itself.

The archive doesn't contain a preprint that explains Diagrammatica's specific error-handling loop—whether the agent revises the specification, queries the backend for clarification, consults its knowledge base, or uses some other strategy to recover from rejection.