LlamaIndex failed function tools: decide Sume retries in the tool

LlamaIndex v0.14.25 changed core to avoid retrying failed function tools, so assume nothing will retry a Sume call for you and make the retry decision inside the tool. Retry only the error categories Sume marks as retryable, always with the same idempotency key, and return a status string for anything else.

The LlamaIndex entry is one line in its release notes ("fix(core): avoid retrying failed function tools", #22841). Sume behavior is from Errors and credits and Jobs and results, read 2026-09-30.

What does a raised exception cost you now?

The release note gives only the headline, so check the source for exact behavior. One plausible reading is that a failed paid tool call is not silently tried again, so a transient queue error would end the tool call unless you handle it yourself.

Which Sume errors are safe to retry?

What about a local timeout?

A timeout in your process is not a failed job. The docs say not to resubmit the original paid request just because a local process timed out; poll with backoff until completed, failed or canceled. Reuse the same key only for the same operation and payload.

What should the tool return?

Catch the error inside the tool and return text the agent can act on, such as "queued, job_id=job_123, poll status" or "validation error: fix the prompt". Keep the job_id in the message so the agent polls instead of resubmitting. A working image-tool example is in LlamaIndex image generation tool with Sume.