Voicelab first-cause: frontend knowledge, backend research, one conversation

A replay of a real conversation (prd, 2026-08-26 evening, stream /agents/voice/2608261852 on the iterate project — read-only evidence): first a question the voice can answer from its own weights, then one that needs the backend and the internet. The original run surfaced the ask()-lane reply losses this eval now guards against: the backend's results arrived as an unsolicited chat message and vanished, leaving the voice saying "the status says results were delivered but I never received them".

Setup#

Same harness as evals/voicelab-roundtrip (read it first): project slug voicelab-eval, everything from apps/os inside doppler run --config prd -- …, utterances synthesized with say -o <file> --data-format=LEI16@16000 --channels=1 "<text>".

Utterances, in this order (the driver plays them sorted, cycling):

1. 01-first-cause.wav: "Hi. Can you explain the first cause argument for God, briefly?"
2. 02-youtube.wav: "Interesting. Now can you find me some recent YouTube debates about that argument?"
3. 03-nudge.wav: "Any luck with those debates yet?"
4. 04-nudge.wav: "Take your time. Anything new come in?"

Run on a fresh timestamped stream:

doppler run --config prd -- pnpm cli voicelab talk --project voicelab-eval \
  --stream-path /agents/voice/eval-cause-<stamp> --converse 5 \
  --utterance-dir <dir> --pretend-speaker <dir>/speaker.wav

Success criteria (all must hold)#

Read the stream's durable events (pnpm cli voicelab transcript --json for the spoken record; itx run for colleague-status / colleague-note events):

1. Frontend handles knowledge itself: an answer-transcript row after the first question explains the argument (mentions "cause"), and the FIRST colleague-status event on the stream comes at a higher offset than the utterance-transcript of the YouTube question — i.e. no note was sent for the knowledge question.
2. Backend roundtrip: at least one colleague-note event exists (the backend's chat messages now arrive as durable events — solicited or not), and an answer-transcript row after the first colleague-note relays research content (mentions "debate", case-insensitive).
3. Statuses flowed: the stream has a "picking up a note from the frontend" status and at least one lifecycle phase ("writing code" or "running code").
4. Clean run: talk exits 0, zero speaker sequence gaps, exactly one agent/created on the colleague stream (/agents/voice-notes/voice/eval-cause-<stamp>), zero on the voice stream.

Timing note: the backend's web research took ~2 minutes in the original conversation; the 5-minute run with nudge utterances absorbs that. If the run ends before any colleague-note arrives and the statuses show work still mid-flight, rerun once with --converse 7 before calling it a failure.