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.
One unspeech deployment = one configKV entry. The split into two keys
modelled a hypothetical split deployment (operators pointing REST and
WS at different unspeech instances) that nobody actually runs, and the
explanatory comments justifying the split were exactly the migration-
narrative anti-pattern we just banned from source.
Schema:
UNSPEECH_UPSTREAM = {
restBaseURL: string, // required, http(s)://host:port
streaming?: { // optional, only when ws TTS is on
baseURL: string, // ws(s)://host:port/...
keys: [{ id, ciphertext }],
adapterParams: {},
models?: [{ id, name?, description? }],
},
}
Admin slice surface flattens to one `kind: 'unspeech'`:
{ kind: 'unspeech', restBaseURL,
streaming?: { upstreamURL, plaintextKey, keyEntryId? } }
Read-site changes:
- routeTts + listTtsVoices read UNSPEECH_UPSTREAM.restBaseURL via
getOrThrow; absent entry → 503 CONFIG_NOT_SET.
- audio-speech-ws dials UNSPEECH_UPSTREAM.streaming and 1008-closes
with streaming_tts_not_configured when the subtree is absent.
- handleListStreamingVoices reads .restBaseURL directly (no more
ws→http scheme swap) and 503s on missing streaming subtree.
- handleListStreamingTTSModels reads .streaming.models.
- config-sync-subscriber listens for UNSPEECH_UPSTREAM invalidations
instead of UNSPEECH_REST_BASE_URL.
Drive-by: tighten ConfigKVService.getOrThrow return to
`Exclude<ConfigDefinitions[K], undefined>` so call sites stop needing
non-null assertions on optional schema entries — the runtime already
throws, the type now reflects it.
Migration: operators repost a single admin slice to rewrite the merged
entry. Old STREAMING_TTS_UPSTREAM and UNSPEECH_REST_BASE_URL rows
become dead data and can be deleted from configKV; nothing reads them.
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.
The two seed scripts are fully superseded by the new admin endpoint
`POST /api/admin/config/router` — same encryption, same configKV
writes, same `configkv:invalidate` publish, plus auth + audit + body
limits. Keeping both code paths created a drift risk on the AAD label
and the merge semantics.
Doc + test fallout:
- `e2e-llm-router.ts` now points readers to the admin endpoint for
the prerequisite seed step.
- `docs/ai-context/verifications/llm-router.md` and
`streaming-tts.md` get curl-based seed instructions; the 2026-05-15
llm-router evidence stays intact with a note that the script it
used has since been removed.
- The U9 follow-up entry in `llm-router.md` flips from "not shipped"
to "partially shipped" — ETag + HMAC publish are still deferred,
so the `config_write` / `config_invalid_hmac` Grafana panels stay
parked.
- Self-edit on the admin route + `app.ts` docstrings to drop the
earlier "scripts stay as break-glass" wording.
Replaces routine use of `scripts/seed-router-config.ts` and
`scripts/seed-streaming-tts.ts` with `POST /api/admin/config/router`.
Operators can now patch one provider at a time without shelling into
the Railway runner; the seed scripts stay as break-glass tools for
cold-boot and disaster recovery.
The endpoint accepts a discriminated-union slice list (openrouter /
azure / dashscope-cosyvoice / streaming-tts), envelope-encrypts
plaintext keys in-process (never echoed back), and supports
merge/reset modes plus dryRun. Writes go through the existing
configKV + Redis `configkv:invalidate` channel so multi-instance
deployments pick up changes within the pub/sub propagation window.
Guarded by the existing `authGuard + adminGuard` pair
(`ADMIN_EMAILS` allowlist + verified email).
Why:
- Add a real bidirectional streaming TTS path: raw LLM tokens are
forwarded to the upstream model (Volcengine v3 via the unspeech ws
bridge) without client-side segmentation, so the model owns sentence
splitting and audio chunks play as they arrive.
- Move audio endpoints out of /api/v1/openai/. `/audio/voices`,
`/audio/models`, `/audio/voices/streaming` are not real OpenAI public
APIs, and the streaming TTS surface has nothing to do with OpenAI —
keeping them under /openai/ mislabelled the contract.
- Introduce `capabilities.speech.transport` on ProviderDefinition so
future streaming providers (ElevenLabs / Cartesia / OpenAI Realtime)
opt in without touching Stage.vue or the session factory.
- Unify Stage.vue's TTS path through a single StageTtsSession so the
chat-orchestrator hooks no longer branch on provider id.
What:
- apps/server: new ws proxy /api/v1/audio/speech/ws bridges client ↔
unspeech with auth, pre-flight flux check, billing from upstream
session.finished.usage, OTel spans.
- apps/server: audio routes moved from /api/v1/openai/audio/* to
/api/v1/audio/* (hard cutover; 404 sentinel tests added).
- apps/server: new /api/v1/audio/voices/streaming proxy reads voices
from unspeech /api/voices?provider=volcengine.
- apps/server: new STREAMING_TTS_UPSTREAM configKV entry +
scripts/seed-streaming-tts.ts.
- stage-ui: new libs/speech/streaming-pipeline.ts opens one ws per LLM
intent (appendText / finish / cancel + onSentence / onError / onDone).
- stage-ui: new libs/speech/tts-session.ts — StageTtsSession interface
with segmenter and streaming adapters; factory dispatches by
capabilities.speech.transport instead of hard-coded provider id.
- stage-ui: providerOfficialSpeechStreaming with capabilities.speech =
{ transport: 'bidirectional-ws' }; settings page with model/voice
picker + ws-based preview.
- stage-ui: Stage.vue chat hooks collapsed to a single currentSession;
hot-swap watcher cancels mid-session on provider/voice/model change;
unmount cancels and drains playback.
Tests:
- 9 streaming-pipeline tests (happy path / buffered / error / cancel /
truncation)
- 11 tts-session tests (factory branch coverage + adapter contracts)
- 4 audio-speech-ws route tests (forwarding / billing / pre-flight /
config-missing)
- 3 legacy-path 404 sentinels in v1 route tests
- Verification doc updated to reflect automated coverage.
When `mapUpstreamError` produced the final 502/503/504, it only carried
`{triedKeys, triedUpstreams, lastStatusCode}` in `details`. The upstream
response body was `.cancel()`'d on the wire and the network error
message vanished into the catch arm — operators staring at a 502 had no
way to tell "OpenRouter region-blocked us" from "key revoked" from
"DNS failed" without re-probing the upstream by hand.
Now each recorded failure carries the diagnostic snippet:
- chat upstreams read at most 256 bytes of the failed body via a
drain-aware reader before cancelling the rest (socket still returns
to the pool, no fallback-storm pool exhaustion).
- TTS upstreams reuse `errorMessageFrom(err)` — adapters already bake
the status + body snippet into `err.message`, so one field carries
both.
- network / timeout attempts record `errorMessageFromUnknown(err)` so
"attempt-timeout" vs "ECONNRESET" vs "DNS failed" stays
distinguishable.
The collected `UpstreamAttempt[]` is attached to `ApiError.cause`
rather than `details`. SEC-5 (no upstream content in client-facing
response body) still holds — only the server-side logger + OTel pick
the cause up. `app.onError` now logs `{details, cause}` together so a
single log line tells the operator both the contract-level summary
and the actual upstream payload.
Adds a router.test.ts regression covering both shapes (HTTP 401 body
snippet + network ECONNRESET errorMessage), with an explicit assertion
that `details` does NOT contain the body text so SEC-5 doesn't drift.
End-state of the multi-step KTD-5 / KTD-6 / U8 work. The knoway sidecar
is no longer reachable from server code; the router is required at boot
and now owns chat completions, TTS synthesis, and voice catalog listing.
Highlights:
- LLM_ROUTER_MASTER_KEY becomes required; app.ts drops the graceful-
skip branch and the chat fallback fetch path is gone.
- /audio/speech and /audio/voices route through new routeTts /
listTtsVoices entries that reuse the chat key-rotator + per-attempt
timeout + abort propagation.
- DEFAULT_CHAT_MODEL / DEFAULT_TTS_MODEL move from env to configKV so
default-model swaps are hot-reloadable via Pub/Sub.
- GATEWAY_BASE_URL removed from env schema, .env, .env.local, smoke,
verification harness. Redis upstream-voices cache deleted — catalogs
come from in-process adapter JSON.
- routeTts splits adapter error contract by ApiError statusCode:
4xx propagates without fallback; 5xx folds into the network-failure
fallback path. handleTTS wraps billing + span attribute in try/finally
to plug a span leak when ttsMeter.accumulate() throws.
- seed-router-config.ts rewritten with --merge (default) / --reset /
--dry-run modes and env-var key handoff (OPENROUTER_KEY / AZURE_KEY /
DASHSCOPE_KEY) so prod seed flows never put plaintext on the CLI.
Adds DashScope CosyVoice seeding.
Docs (CLAUDE.md, architecture-overview.md, transport-and-routes.md)
reflect the new boundary. verifications/llm-router.md replaces the
overstated "U1-U9 shipped" line with an evidence-vs-pending table.
Tests: full 40-file / 343-case server suite green. New regressions pin
ApiError 4xx → no-fallback, ApiError 5xx → fallback, TTS billing
failure → span closed and error propagated.
- Added a new ObservableGauge for distinct active users to track real active user count, mitigating session row inflation issues.
- Updated the Grafana dashboard to reflect changes, including the removal of redundant WS Connections panel and the addition of new metrics for active sessions and distinct users.
- Improved documentation for verification automation processes, outlining a structured approach to automate verification steps and maintain evidence of tests.
- Added a new PostHog client for capturing server-side business events such as Stripe webhooks and subscription state changes.
- Implemented various tracking functions for pricing funnel steps, character creation, and chat session starts.
- Enhanced the flux meter tests to handle partial charges and report unbilled flux correctly.
- Updated the CharacterDialog and Flux settings pages to track user interactions with analytics events.
- Introduced a mechanism to identify users on PostHog based on authentication state to ensure accurate funnel tracking.
- Added necessary dependencies for PostHog integration in the project.
## Description
Adds optional env **`ADDITIONAL_TRUSTED_ORIGINS`**: comma-separated
browser origins that are trusted for **CORS (`/api/*`)**, **Stripe
return URLs**, **Better Auth `trustedOrigins`**, and **dynamic web OIDC
redirect URIs**.
LAN / non-localhost Capacitor dev (e.g. Pocket + Vite on
`https://10.x:5273`) no longer relies on broad private-IP regex;
operators list exact origins in `.env.local` and restart the API server
after changes.
## Linked Issues
<!-- N/A -->
## Additional Context
Pocket iOS dev workflow: `cap`/`capacitor.config` often points at a LAN
HTTPS origin; without this allowlist the API rejects those
`Origin`/`Referer`/`redirect_uri` bases. Review can stay focused on
**`apps/server/src/libs/env.ts`**,
**`apps/server/src/utils/origin.ts`**, and wiring in **`app.ts`**,
**Stripe**, **auth routes**.
- Moved RateLimitMetrics import path to a more centralized location.
- Introduced a new file for active sessions gauge to track user sessions in the database.
- Updated index.ts to include new metrics and ensure proper initialization of observability metrics.
- Modified various routes and services to utilize the new observability structure.
- Added smoke tests for HTTP and WebSocket metrics to ensure proper metric registration and functionality.
- Enhanced error handling for metrics reading failures to improve observability.
- Added new metrics for email service including send, failures, and duration tracking.
- Introduced rate limit metrics to monitor blocked requests and improve abuse detection.
- Enhanced billing metrics to track credited and unbilled flux, as well as TTS character processing.
- Updated OpenAI and Stripe routes to utilize new metrics for better revenue tracking and rate limiting.
- Implemented a smoke test for OpenTelemetry metrics registration to ensure visibility at startup.
The Redis Stream `billing-events` + `worker` Railway role +
advisory-lock poller layered together didn't actually buy us reliability
— `debitFlux` swallowed XADD failures, leaving the door open to "balance
updated, ledger row never written". Collapse the whole thing back to:
`creditFlux` and `debitFlux` write `flux_transaction` ledger rows inline
within the same DB transaction that mutates `user_flux`, and `(user_id,
request_id)` remains the partial unique index that keeps retries safe.
Concrete changes:
- Inline ledger inserts in `BillingService.{debitFlux, creditFlux,
creditFluxFromStripeCheckout, creditFluxFromInvoice}`; drop `billingMq`
and `publishEvent` plumbing entirely.
- `routes/openai/v1` writes `llm_request_log` synchronously via the
existing `requestLogService`; the duplicate `llm-request-log.ts` service
module is removed.
- `bin/run-worker.ts`, `libs/mq/*`,
`services/billing/billing-events.ts`,
`services/billing/billing-consumer-handler.ts`, and matching tests are
deleted. CLI now exposes only `api`.
- `BILLING_EVENTS_*` env vars and the `DEFAULT_BILLING_EVENTS_STREAM`
helper are dropped; `docker-compose.yml` no longer ships a worker
service.
- `docs/ai-context/{workers-and-runtime, billing-architecture,
redis-boundaries-and-pubsub, data-model-and-state,
architecture-overview, README}.md`, `CLAUDE.md`, and the existing
verification docs are updated to describe the single-process synchronous
pipeline.
Tests: 29 files / 247 cases pass. Production deployments need to drop
the worker Railway service after this lands.
RainbowBird