@stackbone/cli 0.7.1 → 0.9.0

package/CHANGELOG.md

@@ -7,6 +7,158 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.9.0] - 2026-11-10
+
+### Added
+
+- **The query rewrite is a card on Studio's System screen.**
+  `GET /api/internal-prompts` lists the query rewrite that
+  `stackbone.rag.retrieve({ rewrite })` runs, with its built-in prompt and the
+  model it falls back to: the `defaultModel` workspace setting, else
+  `openai/gpt-4o-mini`. Its prompt is edited and published through
+  `/api/prompts` like the other System prompts, and `/api/settings` takes
+  `rewriteModel` and `rewriteModelProviderId` (STA3-90).
+- **Voice calls play background audio.** The box sends each call's
+  `backgroundAudio` voice setting to the voice worker, and the worker, the one
+  `stackbone dev` runs too, plays an ambience under the whole call and a
+  thinking sound while the agent works on a reply: the knowledge search, the
+  wait for the model's first token and any box tool call. Before, only a chime
+  played, and only during tool calls. By default a call plays
+  `office-ambience` at 0.7 and `keyboard-typing` at 0.6. Studio's Voice tab
+  gets a **Background audio** group. Deploy the worker on LiveKit Cloud before
+  the boxes: an older worker refuses every call from a box on this release,
+  while the new worker takes calls from an older box and plays the defaults.
+
+### Changed
+
+- **A long hybrid query finds the chunks that hold several of its words.**
+  `stackbone rag query`, Studio's search dialog and
+  `POST /api/rag/collections/{name}/query` used to need every word of the
+  query in one chunk, so a long query found nothing by its exact terms. A query
+  of up to 3 words still needs all of them. From 4 words on, a chunk needs at
+  least 2 of them, `eur` and greetings or sign-offs (`hola`, `gracias`,
+  `saludo`, `regards`…) aside, and the chunks that hold more words rank first.
+  The response shape does not change. Splitting a text into one search per
+  question is an SDK option, `stackbone.rag.retrieve({ rewrite })`, that the box
+  route does not take (STA3-79).
+- **On an OpenRouter box a decision model rules on the input guardrails.**
+  The content moderation, denied topics and prompt injection checks on a user
+  message (chat, voice and workflow input) are decided by `jev-1.13` in one
+  request to OpenRouter's `/systemone` endpoint. Each of these rules takes a
+  `threshold` from 0 to 1, default 0.7, that a score must reach to stop the
+  message; a lower value stops more. A value outside 0 to 1 fails the check.
+  The agent's reply is still judged by the checking model, which ignores the
+  threshold. When the decision model errors, takes longer than 1.5 seconds or
+  leaves a question unscored, the checking model decides and the box log says
+  why. A message the decision model stops carries the reason
+  `the decision model flagged it`, with no scores. A box on another provider,
+  or with no OpenRouter key, checks as before (STA3-58).
+- **The box's own models follow the model provider when nobody picked one.**
+  The box lists the models its provider serves at boot and after every
+  provider save. The guardrail, eval judge and persona simulator models nobody
+  picked run on their code default while the provider serves it or the listing
+  has not answered, and on the `defaultModel` workspace setting when the
+  provider does not serve it. Studio's System screen shows that default, and the
+  memory card shows the one `memory.add` resolves the same way.
+  `POST /api/rag/collections` with no embedding model saves
+  `openai/text-embedding-3-small` as before, unless the provider does not serve
+  it: then the collection gets the first embedding model the provider serves,
+  probed for its dimension, and a provider that lists no embedding model
+  answers 400 `rag_embedding_model_required`. `stackbone rag collections create`
+  sends no embedding model, so it gets the same answer. An OpenRouter box keeps
+  today's models, and a model someone picked is never replaced (STA3-82).
+
+### Fixed
+
+- **Rerank works from creator code in a box.** `retrieve({ rerank })` and
+  `client.ai.rerank.create` call the model through the box's model provider
+  proxy, which had no `rerank` route and answered 404, so every rerank fell
+  back to search order. The proxy now forwards it to the provider's `/rerank`,
+  and a model pinned to a provider goes to that provider (STA3-79).
+
+## [0.8.0] - 2026-11-06
+
+### Added
+
+- **`stackbone build` and `dev` record each agent key's `modelRouting`.** The
+  model inventory now carries the `modelRouting` literal an agent and each of
+  its subagents write, so Studio's Models block shows the code's provider
+  routing next to the one an operator picks. A value OpenRouter does not take
+  is warned about, with its file and line.
+- **Recurring jobs hides the platform's own timers.** Studio's Recurring jobs
+  screen now shows only the workflow crons by default; intake link polling and
+  the box's sweeps appear when **System entries** is on, as on Catalog, Runs and
+  Sessions. A banner counts the hidden ones that are overdue, not armed or
+  failed, so a stuck intake poll still shows. `GET /api/recurring-jobs`
+  returns `system` on every row; an older box leaves it out and Studio shows
+  the row (STA3-71).
+- `stackbone dev` and `stackbone build` seed the `rag-extract` managed workflow
+  for Studio's document value proposals. A creator can override the default in
+  `workflows/rag-extract.workflow.ts` (STA3-67).
+- **The box places outbound calls for `placeCall`.** A workflow's `placeCall`
+  opens a LiveKit room, dispatches the voice worker and dials the number
+  through the trunk in the box secret `LIVEKIT_SIP_OUTBOUND_TRUNK_ID`, with
+  `LIVEKIT_SIP_CALLER_ID` as the optional caller id. The agent's model is
+  warmed while the phone rings. An answered call is kept until the worker
+  reports its end, then the box starts the `onEnd` workflow once. A 30-second
+  sweep ends a call as `failed` when its room is gone past a grace period or
+  it outlived its call token; the sweep is a system entry on Recurring jobs.
+  Migration 0118 adds the calls table (STA3-56).
+- **Each turn of a call is timed on the run that answered it.** The voice
+  worker reports the latencies of every answered turn, and the box writes them
+  as `voice` steps on that run (**Wait**, then **End of turn**, **Speech to
+  text**, **Box**, **Model first token**, **Text to speech** and **Audio
+  out**), which Studio draws in the run's trace. Migration 0119 stores the
+  latencies and 0120 adds the `voice` step type. `stackbone dev` prints one
+  `Voice turn` line per answered turn with the same numbers (STA3-56).
+- **Hybrid search hits say which search found them.**
+  `POST /api/rag/collections/{name}/query` returns optional `dense_rank` and
+  `lexical_rank` on each hit of a hybrid `text` query: its 1-based position in
+  the search by meaning and in the exact-term search, before fusion. A search
+  that did not find the hit omits its rank; `hybrid: false`, an `embedding`
+  query and an older box return none. `stackbone rag query --json` shows them,
+  the table output does not change, and Studio's search dialog labels a hit
+  `meaning`, `exact` or `both` (STA3-66).
+- **`rag.previousExchanges` sets how much of the call the knowledge search
+  reads.** A voice agent's knowledge search sends the last N exchanges of the
+  call (caller question and the reply they heard), oldest first, before the
+  caller's sentence. It takes 0 to 5 and defaults to 1, which sends the same
+  query as before; 0 searches the sentence alone. Studio's voice settings
+  override the code (STA3-76).
+
+### Changed
+
+- **A voice turn's knowledge search reads the previous exchange.** A follow-up
+  such as "¿y los plazos?" names nothing on its own, so the search now carries
+  the caller's previous searched question and the reply they heard, after the
+  output guardrails. A filler turn, or one that failed before speaking, keeps
+  the exchange that held the context (STA3-64).
+- **A voice agent with a knowledge collection only states facts it can
+  source.** Every turn of such a call adds a rule to the prompt: a price, date,
+  instalment, name or duration comes from the knowledge found for the turn, the
+  agent's instructions, a tool result, or an earlier fact of the call the
+  caller said or the agent sourced; otherwise the agent tells the caller an
+  advisor will confirm it. The rule
+  applies even when the search found nothing, and a turn resumed after a tool
+  approval keeps it, whether decided in Studio, by the timeout sweep or after a
+  box restart. A call without a collection sends the same prompt as before
+  (STA3-64).
+- **Voice knowledge search fetches 5 passages by default instead of 3.** At 3,
+  a question that asks two things often found the answer to only one
+  (STA3-72).
+- **Phone callers get LiveKit's telephony noise model.** Their audio is 8 kHz,
+  and the full-band model muffled it and held turns back. With Deepgram Flux
+  deciding the turn, the worker now reports how long Flux's decision took to
+  reach it as the end of turn, which LiveKit reports as 0 (STA3-56).
+
+### Fixed
+
+- **A reply blocked by an output guardrail on a voice call speaks the refusal
+  phrase.** The caller used to hear the guardrail's reason read aloud. The
+  call's `refusalPhrase`, the same one an input guardrail speaks, now replaces
+  it; text chat and the run's step keep the reason. A resume with no live call
+  keeps the reason, since nobody hears it (STA3-59).
+
 ## [0.7.1] - 2026-11-04
 
 ### Added