[prineycom]

Flow e0-stand — epic VOCA-1 «E0 · Стенд: репозиторий, аудио, LLM, Hermes» (research: https://draft.yokeloop.com/artifacts/voca-1-e0-llm-hermes-iim7wlw3). Every subtask was merged --no-ff into flow/e0-stand after review and checks; CI green on each merge.

Tasks

Key	Title	Plan	Review	Checks	Live acceptance
VOCA-2	Репозиторий и запуск (0.1)	si4aeqio	approved (2 rounds)	AGENTS.md Checks green, CI success	local + domain: page, token, /rtc/validate 200
VOCA-3	Hermes preflight — G7 (0.4)	opc6mj2i	approved (2 rounds)	green, CI success	G7 on rpi: passed by engineer decision
VOCA-4	Стенд LLM — G3, G4 (0.3)	zchcvvzz	approved (2 rounds)	green, CI success	G4 passed (TTFT 137.7 ms @8k, 71.1 tok/s); G3 not passed (12/20, 6/10; two-turn 17/20, 9/10)
VOCA-5	Стенд аудио — G1, G2 (0.2)	ldk12j4l	clean (2 rounds)	green, CI success	G1 passed (q8_0); G2 passed with condition (numbers as words + ruaccent stress); end-to-end path measured
VOCA-7	Стенд LFM3-1.7B на кейсах E0	t9d0hecd	approved (2 rounds)	green, CI success	LLM profiles (VOCA_LLM_PROFILE), llama.cpp b12167 default; LFM3-1.7B: neither turn (numbers 7–8/18)
VOCA-8	Стенд LFM2.5-2.6B на кейсах E0	gl3qou1o	approved (2 rounds)	green, CI success	LFM2.5-2.6B: neither turn (numbers 11/18, first sound p50 3.4 s)
VOCA-9	Переключить LLM агента на LFM2.5-2.6B	0a4y23qo	approved (1 round)	green, CI success	VOCA_LLM_PROFILE default lfm2.5-2.6b (bolt-e4b, lfm3-1.7b alternatives); clean clone: fetch --check 101 OK, /v1/models → lfm2.5-2.6b, explicit bolt-e4b → bolt-e4b; stand llm now LFM2.5-2.6B on b12167 (2006 MiB), Nemotron stopped
VOCA-10	Документация E0: b12167 в отчёте G3/G4 и mise trust в README	card only (doc fix, no derive plan)	no separate review (two doc lines, lead)	green	not required

VOCA-6 (LFM3-Omni-2B) was cancelled by the engineer: «я передумал по поводу этой модели, она без ризонинга и русский не поддерживает. Нет смысла ее юзать. Отменяем эту задачу тестируем другие модели».

Epic acceptance: 14/14 PASS on a1f83cb from a clean clone (fetch --check 101 OK, scripts/up.sh all four services, health, token accepted by livekit, https://voca.priney.com/ and /rtc/validate 200, busy-port refusal, down.sh, three reports with G1/G2/G3/G4/G7, six first-hour checks, docs vs host, DNS/Traefik/Hermes). Record: home/projects/yokeloop/voca/tasks/VOCA-1-e0-stand/acceptance.md (mypi home). VOCA-9 was added after the epic acceptance; it changes only the LLM default and docs, and its own clean-clone check covers the llm service. Project checks re-run on the assembled head: green.

Engineer's decisions taken on the way

• Layout/infra: downloads in ~/Apps/~/models, .local/ only for VOCA state; CI on flow/**; domain voca.priney.com via Dokploy Traefik on rpi, DNS-only A records; Hermes API at hermes.priney.com; rtc check «как реальный клиент»; «vite постарше, без файла».
• G7: «Принимать cancelled», «Fallback — норма».
• LLM: «Только клиент» (kwargs), «Время клиента» (TTFT); two-turn mechanic (quick reply A without thinking, B with thinking and tools) — «Замерить в VOCA-4 + в E1».
• Audio: G2 «Пройден с условием»; text preparation (numbers as words + stress) checked before E1 with warm models; voice F4, q8_0, 8 steps; «поезд»-type stress errors accepted.
• Model comparison: one LLM in memory chosen by config; sampling per model card; all on b12167 without re-measuring Nemotron; thinking-on only for small models; «пусть пишет числа цифрами как хочешь решим эту проблему иначе потом»; LFM2.5 quant Q4_K_M.
• Agent LLM: LFM2.5-2.6B instead of Nemotron — «Фиксируем что для агента мы берем эту модель вместо немотрон»; fine-tuning on VOCA cases later; VOCA-9 added to this flow: «да давай добавим в текущий поток»; stopping Nemotron on the stand: «Разрешаю мне (рекомендую)». Architecture docs updated in derive: 8kki4ynb v6, eb1id854 v2.
• Mandatory project rule: everything deployed is documented in the repo (README «Воспроизведение с нуля», .env.example, scripts/SHA256SUMS) and in mypi (INDEX.md, deployment.md).

Inputs for E1

• Turn A without thinking (~1.1 s to first sound) carries no facts or numbers; turn B with thinking (~5 s) — tools must be passed.
• TTS latency is dominated by the first sentence length; small models write long Markdown sentences.
• Numbers normalisation for TTS to be solved separately (the «numbers as words» instruction hurts small models).

🤖 Generated with Claude Code