live: heartbeat sync 2026-10-18T17:49:00-07:00 (Refs #11)

agent/models.yml

@@ -254,3 +254,49 @@ providers:
               # prior assistants' <think> blocks from history. Keep them —
               # multi-turn tool traffic on this endpoint is the majority.
               preserve_thinking: true
+
+  # Speech endpoint — the fleet's TTS, NOT a chat provider. `api:
+  # openai-speech` is a runner API (catalog kind `tts`), so this row never
+  # enters the chat picker, the Ctrl+P cycle, or any chat role: it is
+  # reachable only through the `speech` model role and the tts device.
+  # frank-tts (host port 7772) is the OpenAI /v1/audio/speech façade that
+  # resolves a voice BY NAME; its engine (frank-tts-engine, container-internal
+  # 8880, never published) is arm A qwentts.cpp 5916bf7 — measured 0.55-0.67
+  # RTF warm and 86-451 ms to the first audio byte on Dankclaw's CPU, with
+  # mainline llama-tts (TTS_BACKEND=llama-tts-cli) kept as the rollback.
+  # Transport contract, read out of the harness itself: the request body is
+  # {model, input, response_format, voice?, speed?}; `voice` is forwarded ONLY
+  # when explicitly set, and sampleRate/bitRate are REJECTED outright (the
+  # transport throws "openai-speech does not support sampleRate"), so the
+  # codec comes from the output_path suffix and the only tunable that matters
+  # is the voice name — which must be a real pack name (narrator1/2/3); an
+  # unknown name is an engine error, not a fallback.
+  # DELIBERATELY NOT WIRED YET: config.yml `modelRoles.speech` still points at
+  # local/kokoro. Flip it to "frank-tts (dankclaw)/frank-tts" only after the
+  # service is deployed and has answered a real synthesis, because `auth: none`
+  # makes the provider read as AVAILABLE even while the container is down — the
+  # role would then fail instead of falling back to Kokoro. The same flip must
+  # re-key `speech.voice` (now am_fenrir, a Kokoro voice) to a pack name.
+  # SCHEMA LAW, learned the hard way (2026-10-18): provider-level `api` must be
+  # a CHAT transport — the models.yml schema accepts only the chat/judge APIs
+  # there, and naming a runner API (`openai-speech`) at provider level fails the
+  # file's validation, which DISABLES EVERY custom provider: all three fleet
+  # endpoints vanish from the registry, not just this one (`omp models` prints
+  # "models.yml validation failed — custom providers disabled" and lists only
+  # local/kokoro). The runner api belongs on the MODEL entry only, which is what
+  # the block below does — `auth: none` covers the credential requirement, and
+  # `contextWindow`/`maxTokens` stay unset because the check only fires when they
+  # are provided (the bundled Kokoro row is null on both). `omp models --kind tts`
+  # still prints 128K/16K for this row — the registry fills display defaults — and
+  # the speech transport never reads them: its whole request body is the five
+  # fields named above. Verified 2026-10-18 with `omp models --kind tts` (row
+  # present) and `omp models` (frank-tts absent from the chat roster, all three
+  # fleet providers still listed).
+  "frank-tts (dankclaw)":
+    baseUrl: http://192.168.0.143:7772/v1
+    auth: none
+    models:
+      - id: frank-tts
+        name: "frank-tts (dankclaw) — cloned voices"
+        api: openai-speech
+        input: [text]