[anneoneone]

The backend ends a stream socket whose first audio chunk does not arrive within 15 s, instead of keeping it "on air" with silence. This ports the backend half of #825 (merged to main as 89f5076, in prod as sha-1baf52d) to dev/312-interim. The frontend half is a separate PR on the same issue, because a PR has one lane.

Issue: #827

Validation:

• cargo fmt --check: clean.
• cargo test --locked --all-targets: 169 unit and 23 check-config tests pass, 2 of them new. 2 R2 tests are ignored.
• cargo clippy --locked --all-targets: exit 0. All warnings are in older code; none is in a changed line.
• connection_without_audio_is_ended fails on interim's old receive loop: no frame, timeout. This is the same failure as on main.

Implementation and review

Epic: #312
Executor: Claude
Lane: backend
Reviewer: Codex
Leaf IDs: B827.1.1

Claim: #827 (comment). Done by hand at Anton's request. Board Status is not set: the issue is not on board 2, so set_status.py refuses.

Cause (show 48, 2026-10-30). The host's Safari 18.3 cannot record WebM. The recorder failed, but the panel kept the stream socket open with pings. The status said "live" and "recording", while Icecast played silence for half an hour. Every session logged 0 chunks / 0 bytes received.

Change.

• config.rs: stream_first_audio_timeout_secs (env STREAM_FIRST_AUDIO_TIMEOUT_SECS, serde default 15). Same as main.
• stream_ws.rs: while chunks_received == 0, receiver.next() waits with timeout_at(first_audio_deadline). On timeout it sends NO_AUDIO_ERROR and leaves the loop. After the first chunk the loop is unchanged, so later gaps do not trigger it.

What differs from #825.

• stream_ws.rs: the same hunk. It applies cleanly onto interim's generation logic from fix(stream): keep a replaced producer from stopping its successor #818. The timeout leaves the loop like a dropped connection, so interim's stop_own_stream path runs unchanged. The stream stops unless another connection replaced it, and the recording gets the 30 s reconnect grace (defer_finalize). A replaced socket that never sent audio also ends after 15 s and leaves the current stream alone.
• Tests: main's go_live(state, addr, user, query) is connect_producer here, because broadcast_tests.rs already has a go_live. serve_stream_socket gets a serve_stream_socket_with(overrides) variant that adds the overrides on top of the pipe targets.

Tests (backend/src/broadcast_tests.rs, real router and ffmpeg, 1 s limit).

• connection_without_audio_is_ended: pings only → NO_AUDIO_ERROR, then the stream is not active.
• first_audio_chunk_lifts_the_no_audio_deadline: one chunk, 1.5 s wait → a pong, no error, the stream stays active.

GitNexus. handle_stream_socket impact (upstream): LOW, one caller (stream_ws_handler). detect_changes against dev/312-interim says MEDIUM. Its index is from 32839db, so line shifts map the hunks onto neighbours (producer_busy, overlay_font_path). It lists only the stream_ws_handler flows.

Not in this PR.

• backend/.env.example does not list the new variable (same as main).
• Finalizing an empty recording still tries an R2 upload and logs complete_multipart_upload: service error.
• Main-only stream fixes not ported here: fix(stream): keep reconnect backoff going and reset it on success #819 (reconnect backoff, frontend) and ci(stream): test the live chain against Liquidsoap and Icecast #820 (CI live-chain test).

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
live-moafunk-de	Ready	Preview	Nov 2, 2026 10:51pm UTC