fix(discord): keep voice meetings responsive

[joelneleber]

Summary

• reset Discord voice-channel inactivity timers from inbound audio activity, not just join/playback
• move the default transcript path closer to the Discord bot adapter: STT results are posted to the linked side chat directly by the adapter without invoking the agent for every utterance
• make /voice join default to meeting mode: transcribe voice into the side chat without running a full agent/TTS turn for every utterance
• keep legacy every-transcript agent behavior opt-in via HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS=1
• add targeted voice command/timeout coverage

Test Plan

• /home/j/.hermes/hermes-agent/venv/bin/python -m py_compile gateway/run.py plugins/platforms/discord/adapter.py tests/gateway/test_voice_command.py
• custom venv smoke: adapter-side _process_voice_input default transcript-only mode and opt-in agent callback mode
• custom venv smoke: _handle_voice_channel_input default transcript-only mode and opt-in agent-turn mode
• custom venv smoke: _handle_voice_channel_join default meeting message/auto-TTS disabled and opt-in legacy auto-TTS enabled
• custom venv smoke: DiscordAdapter._note_voice_activity throttles timeout resets

Notes

• Full pytest was not available in the current venv (No module named pytest), so I ran compile + targeted smoke checks with the live Hermes venv.
• scripts/brain-context-check.sh --staged was not present in this repo checkout.

Brain-Context: n/a (brain-context-check script unavailable in this checkout)

[liuhao1024]

Double-posting to side chat when transcript agent turns are enabled

When HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS=1, the voice transcript is posted to the side chat twice:

1. Adapter's _process_voice_input calls _post_voice_transcript_to_side_chat (new in this PR)
2. Adapter then calls _voice_input_callback → gateway's _handle_voice_channel_input, which also posts via channel.send (line ~9454 in current gateway/run.py)

When agent turns are disabled (default), only the adapter posts — correct single posting. When enabled, both post — duplicate messages in the side chat.

Fix: Remove the channel.send block from _handle_voice_channel_input since the adapter now handles side-chat posting. The gateway should only handle auth, dedup, and the agent-turn pipeline.

Also note: the adapter's _post_voice_transcript_to_side_chat sanitizes @everyone/@here, but the gateway's channel.send does not — so the duplicate would be unsanitized.

Minor (code duplication): _env_flag, _voice_transcripts_trigger_agent_turns, and _is_duplicate_voice_transcript are all defined in both gateway/run.py and discord/adapter.py. The gateway versions are now largely redundant since the adapter handles dedup and posting. Consider removing the gateway copies to avoid drift.

[liuhao1024]

Double-posting when HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS=1

When agent turns are enabled, the voice transcript gets posted to the side chat twice:

1. Adapter (_process_voice_input → _post_voice_transcript_to_side_chat) — posts **[Voice]** <@user>: transcript
2. Gateway callback (_handle_voice_channel_input) — posts the same **[Voice]** <@user>: transcript again before dispatching to the agent

The adapter always posts via _post_voice_transcript_to_side_chat, then conditionally calls _voice_input_callback. When the callback is _handle_voice_channel_input and agent turns are enabled, it also posts — producing duplicate messages in the side chat.

Fix: Remove the posting from one side. Either:

• (a) Remove channel.send from _handle_voice_channel_input and let the adapter own all transcript posting, or
• (b) Remove the _post_voice_transcript_to_side_chat call from the adapter and let the callback handle posting (as it did before this PR)

Option (a) is cleaner since the adapter already owns deduplication (_is_duplicate_voice_transcript) and the new _post_voice_transcript_to_side_chat.

---

Duplicate _env_flag helper

_env_flag is defined identically in both gateway/run.py and plugins/platforms/discord/adapter.py. Similarly, _is_duplicate_voice_transcript and _voice_transcripts_trigger_agent_turns are defined in both the gateway and adapter. This creates drift risk — if one copy is updated (e.g., to handle "0" / "false" / "off"), the other won't be. Consider extracting to a shared utility or having the adapter call through the gateway.

[tonydwb]

Code Review Summary

Verdict: Approved

Makes Discord voice meeting mode the default: raw transcripts are posted to the side chat without triggering a full agent turn. This prevents long meetings from becoming slow as session history grows. Operators can opt back into the old behavior with HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS=1.

• Good documentation in docstrings explaining the meeting mode rationale
• New test covers the default (no callback) behavior
• Backward compatible via env flag
• No security concerns

---

Reviewed by Hermes Agent (cron batch)

[liuhao1024]

Double-posting when HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS=1

The adapter's _process_voice_input now always calls _post_voice_transcript_to_side_chat() (line ~2512 in adapter.py), posting the transcript to the side chat. Then, when agent turns are enabled (env var = 1), it also calls self._voice_input_callback(...) which resolves to GatewayRunner._handle_voice_channel_input in run.py — and that handler also posts to the side chat (the channel.send(...) block near the top of the method).

So when the opt-in env var is set, every voice transcript produces two side-chat messages:

1. adapter._post_voice_transcript_to_side_chat() → channel.send(...)
2. run.py._handle_voice_channel_input() → channel.send(...)

Fix: the adapter should skip its own _post_voice_transcript_to_side_chat call when self._voice_transcripts_trigger_agent_turns() is true, since the callback path already handles posting. Or alternatively, _handle_voice_channel_input should skip its posting when called via the callback (e.g. by checking whether the adapter already posted).

---

Minor: duplicated helpers across adapter and gateway

Three helpers are defined identically in both gateway/run.py and plugins/platforms/discord/adapter.py:

• _env_flag()
• _is_duplicate_voice_transcript()
• _voice_transcripts_trigger_agent_turns()

The adapter should own this logic since it lives in the plugin. The gateway should either call through the adapter or import from a shared utility. Independent dedup stores also mean a transcript suppressed as duplicate by the adapter may pass the gateway's independent dedup check (or vice versa).

[teknium1]

Thanks for targeting a real Discord VC inactivity gap: current main only re-arms the timeout after check_silence() returns a completed utterance (plugins/platforms/discord/adapter.py:3162-3179).

Problems

• The opt-in agent-turn path double-posts transcripts. The added adapter post occurs before the callback, and the callback target still posts at gateway/run.py:13061-13066. This matches @liuhao1024's review comments.
• HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS is a new non-secret behavior setting, contrary to AGENTS.md:102-107. It also changes the documented default full-agent VC flow (website/docs/user-guide/features/voice-mode.md:244,351-365).

Suggested changes

• Keep one transcript-post owner and add an enabled-agent-turn regression test asserting one side-chat send.
• Split/rescope the inbound-activity timeout fix; if meeting mode remains configurable, use documented config.yaml plumbing.

Automated hermes-sweeper review.

[teknium1]

When HERMES_DISCORD_VOICE_TRANSCRIPT_AGENT_TURNS=1, the callback immediately below reaches GatewayRunner._handle_voice_channel_input, which still sends the transcript at gateway/run.py:13061-13066. This post therefore creates two identical side-chat messages; make either the adapter or gateway the single posting owner and cover the enabled path.