Files
agent-estate-wiki/incidents/2026-10-31-conduit-voice-local-mode-not-applying.md
T

5.0 KiB
Raw Blame History


title: Conduit "voice mode: local" never applies — Mac mini speech server's LLM proxy is disabled type: incident status: active created: 2026-11-01 updated: 2026-11-01 verified_on: 2026-11-01 confidence: high tags: [conduit, voice, local-live, mac-mini, speech-server, investigation] sources:

  • session 20261031_193709_40b9c2 (Conduit voice session, started 2026-10-31 19:37:09 AWST)
  • ~/.hermes/conduit-settings.json, ~/.hermes/conduit-calls.json, ~/.hermes/config.yaml
  • live probes of the Mac mini speech server (LLM endpoint, /v1/usage, realtime socket)

Conduit "voice mode: local" never applies

Status: OPEN — investigation in progress. Nothing was changed on the Mac mini speech server.

On 2026-10-31 Anthony reported (from the phone): switching the Conduit app's voice setting to "local" shows a successful connection test, but the change never takes effect. A background job launched from that voice conversation (20261031_193709_40b9c2, started 19:37:09 AWST, 30 messages / 17 tool calls) was still open at the 2026-11-01 00:2x audit.

Why the test succeeds and the change still does nothing

The "test connection" button checks socket reachability only — it never exercises the model path. So it passes while every local turn is still impossible:

# Finding Evidence
1 The Mac mini speech server's LLM proxy is disabled its LLM endpoint answers HTTP 501 — "The LLM proxy is disabled. Start the server with --enable_llm_proxy to enable it."
2 The Mac has never served a local turn /v1/usage → llm_proxy.requests: 0 across 25 observed turns
3 The realtime handshake is not understood the Mac's realtime socket accepts the connection but rejects session.update as an unknown event
4 The app's test button proves nothing about the model path it returns success on socket reachability alone

Findings 1–3 are the strongest single explanation: even with voice.mode = local set on the phone, the Mac cannot call a model back. Finding 4 is why the UI reports success.

voice.mode is phone-side, not host-side

The host never reads a voice.mode key — there is none to read:

  • ~/.hermes/config.yaml → voice holds auto_tts, local_live, speech_server — no mode.
  • ~/.hermes/conduit-settings.json contains no voice.* keys at all; its only device state is devices/<UUID>/mode: small for three devices.
  • The device itself reports voice.mode: "classic".

So "set mode to local and the change does not apply" cannot be diagnosed on the host config — the selector lives in the app, and what the host does with it is a local_live path decision.

A misread that was caught mid-session (worth keeping)

local_live.py:882 is the only writer of _local_live_last_request, and it sits inside the chat-completions route. A "recent last_request_at" reading during the investigation therefore indicated this investigation's own probe, not phone traffic — the agent initially read it as phone activity and corrected the call in the same session. Live check: conduit-calls.json → last_request_at is null.

Lesson: on this integration, last_request_at is proof of some HTTP request hitting the local-live route, not proof of a phone call.

Configuration at the time of writing

Changed during the 2026-10-31 session (config.yaml mtime 21:02:30):

  • voice.local_live.front_model = {base_url: http://100.78.63.20:20128/v1, model: mistral/voxtral-small-latest} — was mistral/ministral-14b-latest.

Unchanged: voice.speech_server.url = wss://…:8443/v1/realtime, voice.auto_tts: false, tts.provider: ollama_local, stt.provider: local.

The Conduit dashboard gateway was restarted at 19:35:02 (pid 176771), minutes before the session opened.

Prompting context (candidate, not proven)

The plugin self-updated to 0.122.2 on 2026-10-31 18:05, which drove a gateway reload at 18:06:41 → incidents/2026-10-31-gateway-reload-force-drain-orphaned-run. A newly shipped local-live path is a reasonable regression candidate, but the disabled LLM proxy would break local mode on any plugin version, so the version change is not yet implicated.

Open items

  • Enable the Mac mini's LLM proxy (--enable_llm_proxy) or confirm the intended Local Live architecture — with the proxy off, finding 1 alone explains the symptom.
  • Confirm the realtime event vocabulary the plugin expects vs. what the Mac's socket implements (session.update is currently rejected as unknown).
  • Decide whether front_model should be a text or audio model — a chat-completions base_url pointed at an audio model may itself be the mismatch.
  • Re-test from the phone after any change; the app's own test button is not evidence.