Track per-app TTS concurrency in Redis, route capped upstreams by
available pool capacity, and surface pool saturation metrics. Document
the Voice Pack plan so the remaining backend and card-binding work has
an explicit implementation map.
Frontend voice picker silently lost Azure voices because /audio/voices
was called without ?model=, so the server fell back to DEFAULT_TTS_MODEL
and only ever returned one model's catalog. /audio/models also hardcoded
a single `auto` alias, so the client could never request a specific
model id even if it wanted to. On top of that, the Azure adapter shipped
only 2 hand-typed voices in voices/azure.json — the rest of Microsoft's
catalog was unreachable.
Fixed in one pass:
- Drop voices/*.json. Every getVoiceCatalog now calls unspeech REST
(backend=microsoft|alibaba|volcengine). Azure proxies Microsoft's live
voices/list; cosyvoice + volcengine use unspeech's embedded catalogs.
- Drop direct upstream provider calls from send(). All three adapters
POST to <UNSPEECH_REST_BASE_URL>/v1/audio/speech with the openai-shaped
body unspeech expects (model: `<backend>/<id>`). Azure SSML still built
on our side so <prosody rate> speed survives — unspeech accepts
pre-built <speak> documents.
- Fail-fast on every voices endpoint: empty ?model= → 400 MISSING_MODEL
(no implicit DEFAULT_TTS_MODEL fallback), only `auto` resolves;
streaming upstream missing → 503; unspeech non-2xx / network err → 502;
malformed body → 502 with grepable message. No silent `{voices: []}`.
- /audio/models lists every LLM_ROUTER_CONFIG.tts.models key + `auto`.
- Frontend providerOfficialSpeech.listVoices passes ?model= and throws
on upstream failure instead of returning [].
- New UNSPEECH_REST_BASE_URL configKV entry (no default, missing → 503).
config-sync subscriber invalidates the voice cache when this key OR
LLM_ROUTER_CONFIG changes.
- Router voice catalog Redis cache covers all providers with per-provider
TTL (azure 6h live, alibaba/volcengine 24h since unspeech embeds them
at build time).
Pending billing-SKU verification: DashScope cosyvoice now goes through
unspeech's WS-internal alibaba backend (was REST `/SpeechSynthesizer`).
Functional output equivalent — confirm bill after deploy, revert via
git history if WS pricing is materially higher.
Why
- src/services/ was an unordered mix of single-file services and module
directories with no shared classification axis, plus several long-dead
admin batch helpers that survived the move to the simpler synchronous
admin-flux-grants flow.
What
- services/ now has two top-level layers:
domain/ — DB state + business rules (billing, characters, chats,
flux, flux-transaction, llm-router, providers, request-log,
stripe, user-deletion, admin/{flux-grants,router-config})
adapters/ — thin wrappers over external SDKs / infra (config-kv, email,
posthog, tts/)
- admin/* moved under domain/admin/ with consistent plural names
(flux-grants, router-config).
- tts-adapters/ collapsed to adapters/tts/ (no redundant -adapters suffix
once nested under adapters/).
- 63 src files + scripts/e2e-llm-router.ts + tests/verifications/_harness.ts
had relative imports rewritten; git mv preserves blame.
- apps/server/CLAUDE.md and docs/ai-context/*.md updated to match new paths.
Dead code removed
- services/admin-flux-grant-batches/ (service + worker + tests, 1090 LOC) —
superseded by admin-flux-grants and never wired into app.ts.
- routes/admin/flux-grant-batches/ — same.
- utils/redis-compressed.ts + test — zero production call sites.
- llm-router/index.ts re-exports trimmed from 26 to 6; only symbols with
external consumers are kept.
Intentionally kept
- schemas/flux-grant-batch.ts and its schemas/index.ts export remain so the
drizzle-kit generate diff stays empty. Removing them is a separate PR
that owns the drop-table migration for flux_grant_batch /
flux_grant_batch_recipient.
Verification
- pnpm -F @proj-airi/server typecheck: passes.
- pnpm exec eslint apps/server: 49 errors, identical to main baseline
(all are pre-existing node/prefer-global/buffer in envelope-crypto and
scripts/e2e-llm-router; untouched by this change).
- Vitest passes per-file; the 6 mockDB hook timeouts under full-parallel
run are the known pushSchema-per-worker infra cost, not a regression.
RainbowBird