Skip to content

fix: keep sarvam stt streams open across flushes - #7990

Merged
chenghao-mou merged 2 commits into
mainfrom
chenghao/fix/AGT-3516-sarvam-stt-flush
Oct 20, 2026
Merged

chenghao-mou merged 2 commits into
mainfrom
chenghao/fix/AGT-3516-sarvam-stt-flush

Conversation

@chenghao-mou

Copy link
Copy Markdown
Member

Extracted from #7989. Since #3498, legacy Sarvam STT has treated flush() as terminal, discarding buffered audio and stopping subsequent audio sends. Flush now drains the buffer and sends a provider flush when enabled. Only end_input() sends end_of_stream.

Framework manual_flush opt-in stays in #7989. Verified with fake WebSockets; live Sarvam behavior remains unverified.

Addresses AGT-3516

Initial prompt and agent context

Model: GPT-6 Astra

but the sarvam change is more like a bug fix?

yeah, create a new PR for that fix. Can you trace how it ended up like that in the first place?

@chenghao-mou
chenghao-mou marked this pull request as ready for review October 20, 2026 16:40
@chenghao-mou
chenghao-mou requested a review from a team as a code owner October 20, 2026 16:40
devin-ai-integration[bot]

This comment was marked as resolved.

Server VAD already finalizes its segment. Echoing a flush on the next input frame can prematurely finalize the next utterance. Remove that deferred flush and cover delayed END_SPEECH events.

Verify repeated flushes and automatic VAD against live Saaras v3 and v4 on one WebSocket.
@chenghao-mou
chenghao-mou merged commit 5f81058 into main Oct 20, 2026
25 checks passed
@chenghao-mou
chenghao-mou deleted the chenghao/fix/AGT-3516-sarvam-stt-flush branch October 20, 2026 18:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants