@stackbone/cli 0.9.2 → 0.9.4

package/CHANGELOG.md

@@ -7,6 +7,73 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.9.4] - 2026-11-12
+
+### Added
+
+- **LiveKit Inference as a model provider.** Once the LiveKit credentials are
+  saved under Voice, the box lists a **LiveKit Inference** provider with the
+  LLMs LiveKit hosts (Gemma 4 31B, Gemini, GPT-4.1 and GPT-5.x among them).
+  Pick one of its models on an agent in Studio: the box signs each call with
+  the Voice credentials, retries a 429 or 503 twice, and names the model when
+  LiveKit refuses it for the project's data region. Its models run only where
+  they are picked from it, so no other agent changes provider. The provider
+  cannot be added, edited or removed by hand (PLA-2416, PLA-2417).
+- **Test connection on LiveKit Inference** sends a one-token request signed
+  with the Voice credentials (PLA-2417).
+
+### Fixed
+
+- **The browser tools that use a model work behind the box's model provider
+  proxy** and use the model `browserTools({ model })` names (STA3-120).
+- **A pinned model call records its vendor.** A call on
+  `livekit::google/gemma-4-31b-it` is recorded with the provider `google`,
+  not `livekit::google` (PLA-2416).
+- **Test on a webhook trigger with no event yet gives no verdict.** The
+  declared sample is another account's example, so a correct mapping no longer
+  reads red before the link has its first event. A key or a destination schema
+  that does not compile is still refused (STA3-121).
+
+Studio shows the LiveKit Inference provider from this release on. Deploy
+Studio before a box runs this version: an older Studio cannot read the
+provider list it answers.
+
+## [0.9.3] - 2026-11-12
+
+### Added
+
+- **Call variables in the voice greeting.** The box fills the greeting's
+  `{{name}}` placeholders from `placeCall({ variables })` or the variables a
+  Studio test call passes, when it opens the room, and gives every model call
+  of the call the same variables as data. A missing variable is left out and
+  the punctuation around it tidied (PLA-2409).
+- **New voice settings** `speakFirstClause`, `noiseCancellation`,
+  `interruptions.mode` and `timeZone`, in the agent's code and in Studio's
+  Voice tab. A call's date and knowledge now ride the caller's last message,
+  not the system message, so a provider can cache the call's prompt.
+- **A written turn to a voice agent searches its knowledge** the way a call
+  turn does: Studio chat, the API, a connected channel and eval cases
+  (STA3-104).
+- **Guardrails on replies:** a voice call's reply is judged sentence by
+  sentence and the caller's heard text is kept after a cut (STA3-60), the
+  decision model judges replies too (STA3-101), and a tool never runs behind a
+  blocked reply or holding line (STA3-102).
+- **A webhook trigger link is not judged until its first event** (STA3-112).
+
+- **The thinking sound waits for the holding line.** When a voice agent says a
+  line right before a box tool, such as "Let me check that", the voice worker,
+  the one `stackbone dev` runs too, starts the thinking sound 300 ms after the
+  line has played out, so the typing no longer plays under it. It starts at
+  once when the line's speech fails, and at most the new
+  `backgroundAudio.holdingLineMaxWaitMs` voice setting (0 to 10000 ms,
+  default 10000) after the tool start. On an agent with an output guardrail,
+  the box judges and releases the line when the tool starts, so the caller
+  hears it while the tool runs (STA3-99).
+- **Studio's Voice tab gets a Holding line max wait (ms) field**, after
+  **Thinking delay (ms)**. An older box that sends no wait gets 10000 from the
+  voice worker. Deploy the worker on LiveKit Cloud before the boxes: an older
+  worker refuses a call whose settings carry the wait.
+
 ## [0.9.2] - 2026-11-10
 
 ### Added