@klex/agent-admin-api 0.11.0 → 0.11.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +2 -2
- package/dist/index.d.ts +695 -707
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -36,11 +36,11 @@ clients must not split model IDs on colons.
|
|
|
36
36
|
|
|
37
37
|
## Realtime voice client contract
|
|
38
38
|
|
|
39
|
-
- `GET /v1/providers/{id}/voices?modelId=...&refresh=false` returns `
|
|
39
|
+
- `GET /v1/providers/{id}/voices?modelId=...&refresh=false` returns `voices`, optional `defaultVoiceId`, and optional `stale`. Each voice contains opaque `voiceId`, `displayName`, optional `description`, and `isPreviewAvailable`. An empty list means no selectable voices are available for this provider/model. Provider capabilities advertise discovery/preview hooks. Listing does not generate audio. Encode the complete model ID in the query, including slashes/colons.
|
|
40
40
|
- `POST /v1/providers/{id}/voices/preview` takes `{ modelId, voiceId }`. It can trigger billable same-model Live generation and returns binary `audio/wav`, not JSON/base64 or a public URL. `X-Voice-Preview-Model` is URI-encoded and `X-Voice-Preview-Mode` is `live`. On failure parse the JSON error contract; 429 means quota, 502 means upstream generation failure. Arbitrary text/URLs/model overrides are not accepted.
|
|
41
41
|
- `PATCH /v1/settings/model-selection` saves `voice.sts` references with optional `voiceId`. Preserve both `voice.tts` and `voice.stt` because the voice object is replaced. Omit `voiceId` in the replacement STS reference to reset. Other selection purposes reject it. New calls use changes; active calls do not switch voices.
|
|
42
42
|
|
|
43
|
-
Existing omitted defaults remain: OpenAI Realtime/GPT-Live `marin`, Gemini's provider default. Unknown/custom model profiles remain usable with defaults, but have no selectable catalog.
|
|
43
|
+
Existing omitted defaults remain: OpenAI Realtime/GPT-Live `marin`, Gemini's provider default. Unknown/custom model profiles remain usable with defaults, but have no selectable catalog. `defaultVoiceId` identifies a known default for previewing without saving an explicit choice. `stale: true` means discovery failed and the last successful cached catalog was returned; ordinary cache hits do not set it. Cold-start discovery failures return errors. Samples improvise a short introduction from the current agent name and full soul. Realtime/Gemini generate it directly; GPT-Live uses the first configured chat model to write commentary before speaking it with the selected Live voice. That extra text call is potentially billable and requires a chat selection. GPT-Live may paraphrase commentary. Name/soul changes bypass old cached samples; different voices may use different wording. Browser clients fetch over their existing authorized Cloud path, check `response.ok`, then use `response.blob()` and an object URL for playback (revoke it afterward). Headless/TUI clients use `arrayBuffer()` and an external WAV player; the backend does not implement playback.
|
|
44
44
|
|
|
45
45
|
Local admin access remains unauthenticated and loopback-only; Cloud access remains authorized through the existing tunnel. Responses use private/no-store headers, while the agent caches successful samples in bounded memory for reuse. Previewing never starts a production call or supplies history, tools, or credentials as prompt context. It intentionally supplies the full soul to the speech provider or, for GPT-Live, the configured chat provider. Prompts instruct models not to speak private details, but generated output still depends on provider behavior. No UI selector is included in this contract change.
|
|
46
46
|
|