GitHub user da-daken added a comment to the discussion: Parallel Tool Call
Execution
Thanks for the quick review, glad those three are closed. Two things:
1. **Default** — I'd still keep it on by default. You're right that the exposed
users are the ones who don't read docs, but they won't flip the opt-in switch
either, so opt-in mostly just denies everyone the latency win while those users
stay exposed anyway. Default-on gives everyone the speedup, and the ones who
actually hit excessive duplicate calls get a concrete symptom that leads them
to the docs. The exposure is also narrow (only side-effecting tools without a
reconciler). I'd rather surface the risk where they'll see it — a prominent
note in the agent-authoring tooling/skill — than hide the whole feature behind
opt-in. If it turns out worse in practice, flipping back to false is a
one-liner.
2. **Timeout round-trip** — good catch, genuine oversight. I'll leave it as-is
for v1: failures and timeouts go through the exact same path today (both into
the error map, nothing branches on "was it a timeout"), so the recovered-status
divergence has no observable effect. I'll fix the doc's over-claim ("rebuilt
into TIMED_OUT by exception type on recovery") and mark it a known v1
limitation. When we actually need to branch on timeouts (e.g. a dedicated retry
policy), we'll promote TIMED_OUT to a real persisted status and separate it
cleanly then.
GitHub link:
https://github.com/apache/flink-agents/discussions/855#discussioncomment-17647761
----
This is an automatically sent email for [email protected].
To unsubscribe, please send an email to: [email protected]
