Skip to content

feat(desktop): hang up GPT-Live after idle timeout - #111102

Open
1MORLAP wants to merge 1 commit into
NousResearch:mainfrom
1MORLAP:feat/gpt-live-idle-hangup
Open

feat(desktop): hang up GPT-Live after idle timeout#111102
1MORLAP wants to merge 1 commit into
NousResearch:mainfrom
1MORLAP:feat/gpt-live-idle-hangup

Conversation

@1MORLAP

@1MORLAP 1MORLAP commented Sep 14, 2026

Copy link
Copy Markdown

Summary

  • Desktop GPT-Live had no inactivity hangup, so a forgotten live call keeps billing OpenAI at $0.05/min until stop/quit/disconnect.
  • Default 300s of silence (voice.gpt_live.idle_hangup_seconds; 0 disables). Settings → Voice row included.
  • Timer pauses while a Hermes delegation is in flight (KeyArgo on Desktop GPT-Live: default idle hangup so forgotten sessions stop billing #111075): long jobs must not look like silence.
  • Chat session and in-flight work are not cancelled; only the WebRTC live session closes.

Closes #111075

Test plan

  • scripts/run_tests.sh tests/tui_gateway/test_voice_live_delegation.py (8 passed, including idle parse/default/zero)
  • apps/desktop vitest src/lib/voice-live-idle.test.ts (policy: deadline, 0=off, pause on delegation)
  • Settings → Voice shows GPT-Live Idle Hangup; set 10, start GPT-Live, stay quiet → hangup notice; set 0 → no hangup
  • Ask a long Hermes job, stay quiet during it → must not hang up until the turn finishes plus idle

Does not overlap #79574 / #76889 (chained mic only).

@1MORLAP
feat(desktop): hang up GPT-Live after idle timeout
cc1a414 
Forgotten full-duplex sessions keep billing OpenAI at $0.05/min.
Default 5 quiet minutes (voice.gpt_live.idle_hangup_seconds, 0 disables).
The timer pauses while a Hermes delegation is in flight so a long job
cannot look like silence.

Closes NousResearch#111075
@alt-glitch alt-glitch added type/feature New feature or request P2 Medium — degraded but workaround exists comp/desktop Electron desktop app (apps/desktop/*) tool/tts Text-to-speech and transcription area/usage-cost Token accounting, usage reporting, billing, cost tracking area/config Config system, migrations, profiles labels Sep 14, 2026
@kvnloo

kvnloo commented Sep 15, 2026

Copy link
Copy Markdown
Contributor

Comment assumes head cc1a414db19de1afa65f09b428f70b59caf1e014; void if moved.

KEEP — Desktop GPT-Live idle hangup (cost safety)

  • Default voice.gpt_live.idle_hangup_seconds: 300 (0 disables), Settings → Voice row, and hangup that closes only the WebRTC live session (chat + in-flight Hermes work stay) is the right product safety net for forgotten GPT-Live billing ($0.05/min idle counts).
  • Pausing the deadline while a Hermes delegation is in flight matches the KeyArgo scope note on Desktop GPT-Live: default idle hangup so forgotten sessions stop billing #111075 — silence during a long turn must not look like walk-away idle.
  • Prefer this as the idle-lifecycle native GPT-Live vehicle; peer complementary fix is feat(voice): add Codex subscription billing to native GPT-Live #108940 (Codex subscription auth on the same live surface) — neither replaces the other (shared files, different axes).

CHECK

One-click close / ordering

Landing-evidence

  • When merged, please reply with the merge commit SHA on main (not “completed”).

This was referenced Sep 15, 2026
Open
Open
Open
Closed
Open
@kvnloo

kvnloo commented Sep 15, 2026

Copy link
Copy Markdown
Contributor

Comment assumes head cc1a414db19de1afa65f09b428f70b59caf1e014; void if moved.

KEEP — Desktop GPT-Live idle hangup (cost safety)

  • Default voice.gpt_live.idle_hangup_seconds: 300 (0 disables), Settings → Voice row, and hangup that closes only the WebRTC live session (chat + in-flight Hermes work stay) is the right product safety net for forgotten GPT-Live billing ($0.05/min idle counts).
  • Pausing the deadline while a Hermes delegation is in flight matches the KeyArgo scope note on Desktop GPT-Live: default idle hangup so forgotten sessions stop billing #111075 — silence during a long turn must not look like walk-away idle.
  • Prefer this as the idle-lifecycle native GPT-Live vehicle; peer complementary fix is feat(voice): add Codex subscription billing to native GPT-Live #108940 (Codex subscription auth on the same live surface) — neither replaces the other (shared files, different axes).

CHECK

One-click close / ordering

Landing-evidence

  • When merged, please reply with the merge commit SHA on main (not “completed”).

This was referenced Sep 15, 2026
Open
Open
Open
Open
Open
Open
Open
@alt-glitch alt-glitch mentioned this pull request Sep 15, 2026
Closed
@Finn763

Finn763 commented Sep 15, 2026

Copy link
Copy Markdown
Contributor

Design deltas from #111611 (I'm closing that PR as a duplicate of this one — same voice.gpt_live.idle_hangup_seconds, same 300 s default, same Settings row).

Two things in mine that might be worth taking if they aren't already here:

1. Injectable clock (testability). The watchdog takes its time source as an option instead of calling Date.now() directly:

new VoiceLiveSession(handlers, { idleHangupSeconds, now })   // now: () => number, default Date.now

so the idle deadline is asserted with a fake clock in unit tests (threshold-1 does not fire / threshold fires, 0 never fires over a simulated hour, a config of 45 is honoured) rather than with real timers.

2. Per-turn activity, not just audio. Idle is reset by every one of: vendor transcript deltas (either speaker), session.started, session.delegation.created, loud frames in the existing armSpeakingProbe, every outbound event (commentary/thinking/instructions/mute) — plus a public noteActivity() the composer hook calls on each 200 ms reply tick while a Hermes turn is in flight. That last one is the difference between "the model is thinking and the user is silent" counting as idle or as busy; without it a long tool-running turn can be hung up mid-work.

3. Poll tick vs re-armed timeout. The deadline is polled on a 1 s interval armed at construction (cleared in finish()), so the injected clock stays the only time source and a mic-permission prompt left open still hits the deadline. Teardown sends exactly {type:'session.close'} and reports onClosed('idle_timeout'), so billing stops at the close event rather than at the vendor's echo (which is why a manual stop still reports close_requested).

Not a claim that mine is better overall — take whatever is useful, ignore the rest.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

area/config Config system, migrations, profiles area/usage-cost Token accounting, usage reporting, billing, cost tracking comp/desktop Electron desktop app (apps/desktop/*) P2 Medium — degraded but workaround exists tool/tts Text-to-speech and transcription type/feature New feature or request

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

Desktop GPT-Live: default idle hangup so forgotten sessions stop billing

4 participants

@1MORLAP @kvnloo @Finn763 @alt-glitch