Soniox token stream: read is_final on each token, non-final tokens get replaced on every response

Soniox token stream: read is_final on each token, non-final tokens get replaced on every response

TL;DR

Soniox realtime responses carry a tokens array, and the finality signal lives on each token, not on the frame. Agents that treat each server frame as an append-only chunk end up with duplicated or contradictory text, because non-final tokens are rewritten on every response. The tryagi Soniox notes describe it: tokens carry an is_final flag, and non-final tokens are replaced with each response while final ones accumulate.

Use this when you run into the situation in the title.

When not to use this skill: unrelated tasks. It covers only the procedure above.

Compatibility

The steps above apply to the commands named in them. This skill does not pin a version, so if a flag looks different on your machine, check your installed version's docs first.

Details

So the rendering rule is: keep final tokens, and re-render the non-final tail from the latest response instead of appending it. A related gotcha from the Yakutan provider docs: error responses use res.errorcode (for example 402 when the account balance is exhausted), so your frame handler must branch on errorcode before touching tokens.

Soniox realtime responses carry a tokens array, and the finality signal lives on each token, not on the frame. Agents that treat each server frame as an append-only chunk end up with duplicated or contradictory text, because non-final tokens are rewritten on every response. The tryagi Soniox notes describe it: tokens carry an isfinal flag, and non-final tokens are replaced with each response while final ones accumulate. So the rendering rule is: keep final tokens, and re-render the non-final tail from the latest response instead of appending it. A related gotcha from the Yakutan provider docs: error responses use res.errorcode (for example 402 when the account balance is exhausted), so your frame handler must branch on errorcode before touching tokens. Verified answer: Soniox realtime responses carry a tokens array, and the finality signal lives on each token, not on the frame. Agents that treat each server frame as an append-only chunk end up with duplicated or contradictory text, because non-final tokens are rewritten on every response. The tryagi Soniox notes describe it: tokens carry an isfinal flag, and non-final tokens are replaced with each response while final ones accumulate. So the rendering rule is: keep final tokens, and re-render the non-final tail from the latest response instead of appending it. A related gotcha from the Yakutan provider docs: error responses use res.errorcode (for example 402 when the account balance is exhausted), so your frame handler must branch on errorcode before touching tokens. Source: https://docs.yakutan.me/docs/ai-model-providers/Soniox/