petri-retry-patterns

To retry only the failed worker in a fan-out parallel review workflow, attach per-worker RetryPolicy configurations or explicit retry transitions with Trial counters directly to individual worker failure paths, avoiding downstream rerun-all loops.

A downstream Verdict-triggered rerun-all retry loop breaks with an AND-merge because it forces already successful workers to re-execute, causing the fan-out parallel review workflow to get stuck or redundantly process completed tasks.

Correct Petri net wiring for per-worker retry policies in LLM orchestration requires selecting transition-level RetryPolicy per worker or an explicit per-worker retry transition with Trial counters, while strictly avoiding any retry transition placed after Verdict.

Separate transient worker execution failures from semantic review failures by distinguishing individual worker execution retries from downstream merged result repairs, ensuring retry triggers attach to specific worker failure paths rather than final verdicts.

Before generating a fan-out Petri net with workflow retry policies, validate RetryPolicy usage, verify guards, check GiveUp transitions, and ensure trial counter correctness to prevent incorrect wiring that reruns successful workers.

You should not use a downstream rerun-all retry design in parallel review workflows when multiple workers contribute to an AND-merge, as this anti-pattern causes the workflow to get stuck by attempting to rerun already successful workers.