turbollm 1.14.1 → 1.14.3

package/README.md

@@ -249,7 +249,10 @@ the detail:
   **prefill t/s**, **time-to-first-token**, **total time**, exact **token counts**, and a
   **context-usage meter** (filled / max) on every reply.
 - **Thinking control** — toggle reasoning **off** for a direct answer, or leave it **on** with
-  collapsible, timed "thought for N s" blocks.
+  collapsible, timed "thought for N s" blocks that grow with the reasoning; drag a block's bottom
+  edge to size it once and every block remembers it.
+- **Find in chat** — **Ctrl/Cmd+F** (or the header's search button) searches the open conversation:
+  "3 of 17", Enter or the up/down buttons jump between matches and pan the chat to each, Esc closes.
 - **Markdown + syntax-highlighted code** with one-click copy — plus **inline Unicode charts**
   the model draws when a comparison, trend, or hierarchy is genuinely worth a visual.
 - **Live artifacts** — `html`, `svg`, and `mermaid` replies render as **sandboxed, offline
@@ -369,14 +372,20 @@ there. Entirely local; nothing leaves your machine.
 </details>
 
 <details>
-<summary><strong>🧠 Jev — ask structured questions about any text, locally</strong></summary>
+<summary><strong>🧠 Text classification (Jev and Laya) — ask structured questions about any text, locally</strong></summary>
 
 <br/>
 
+- **Models → Discover → Text classification** finds every model TurboLLM can run for this, whichever chat
+  engine is active. Two kinds run today: Jev (NLI cross-encoders on vLLM) and Laya (its own engine).
 - **Load an NLI model (such as OpenJev) and TurboLLM detects it as a Jev model.** Workspace turns
-  into a JSON-first playground: two editors — the content, and the questions you're asking about
-  it — *are* the request. Ctrl/Cmd+Enter to run; see the answers per question, the raw response,
-  and a copyable `curl`.
+  into a playground: two editors — the content, and the questions you're asking about it — *are*
+  the request. Ctrl/Cmd+Enter to run; see the answers per question, the raw response, and a
+  copyable `curl`.
+- **The questions editor has a Form view, too.** Toggle between JSON and Form any time. Form gives
+  you one card per question — a type dropdown (Yes/no, Pick one, Scale), instructions, and the
+  fields each type needs (an option list, an ordered scale) — no hand-written JSON required. Both
+  views stay in sync, and editing one question never touches another's.
 - **`POST /v1/systemone`** answers yes/no, pick-one, and scale questions in one call, using the
   public System One API's exact field names — a client written for that shape only changes its
   base URL and key. `POST /v1/classify` and `POST /v1/rerank` are there too for raw
@@ -387,7 +396,8 @@ there. Entirely local; nothing leaves your machine.
 - **Laya decision models, too** ([convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya)):
   - What it is: a calibrated System One model that answers the same `/v1/systemone` questions in 100+ languages.
   - Where it runs: on its own Laya engine (Engines page; Windows, macOS, Linux; CPU or GPU), beside your chat model.
-  - How to get it: search "laya" in Discover. The playground opens from the Workspace sidebar while it is loaded.
+  - How to get it: switch on the Text classification chip in Discover, or search "laya". The same playground opens from the Text classification tab under Chat, Code and Routines.
+  - Overview of both: <https://turbollm.dev/docs/text-classification>.
 
 </details>
 

package/dist/{chunk-P2FJIT7T.js → chunk-SEL52AIK.js}

@@ -76,6 +76,16 @@ async function proxyStream(remote, path, init, signal, fetchImpl = fetch) {
   return fetchImpl(buildUpstream(remote, path), { ...init, signal });
 }
 
+// src/models/text-classifier.ts
+function isTextClassifier(entry) {
+  return textClassifierRuntime(entry) !== void 0;
+}
+function textClassifierRuntime(entry) {
+  if (entry.jev) return "vllm";
+  if (entry.laya) return "laya";
+  return void 0;
+}
+
 // src/observability/request-log.ts
 import { randomUUID as randomUUID3 } from "crypto";
 function requestLogConfig(d) {
@@ -195,12 +205,13 @@ function resolveChatUpstream(d, requestedModel) {
   if (ms.state !== "running" || !ms.model) {
     return { ok: false, status: 409, code: "model_not_loaded", message: "Load a model first." };
   }
-  if (d.scanner.get(ms.model.key)?.jev) {
+  const loadedEntry = d.scanner.get(ms.model.key);
+  if (loadedEntry && isTextClassifier(loadedEntry)) {
     return {
       ok: false,
       status: 409,
       code: "jev_model_loaded",
-      message: "A Jev model is loaded \u2014 it labels text and cannot chat. Switch to a chat model, or use the Jev Playground."
+      message: "A text classification model is loaded \u2014 it cannot chat. Switch to a chat model, or use the text classification playground."
     };
   }
   const target = d.manager.target();
@@ -1921,7 +1932,7 @@ var CREATE_ROUTINE_TOOL = {
         prompt: { type: "string", description: "The task to run, in the same words a user would ask it." },
         scheduleDisplay: { type: "string", description: 'Human-readable schedule, e.g. "Runs weekdays at 9:00 AM". Must accurately describe scheduleRule.' },
         scheduleRule: SCHEDULE_RULE_SCHEMA,
-        modelKey: { type: "string", description: `One of TurboLLM's own model keys \u2014 a compound id (e.g. "gemma 4 26b a4b qat|Q4_0|14439362752"), never a generic name like "gpt-4" or "claude". Call list_models first if you don't already know a real one; never guess. Never pick a model list_models marks "kind: jev" \u2014 it labels text and cannot run a routine.` },
+        modelKey: { type: "string", description: `One of TurboLLM's own model keys \u2014 a compound id (e.g. "gemma 4 26b a4b qat|Q4_0|14439362752"), never a generic name like "gpt-4" or "claude". Call list_models first if you don't already know a real one; never guess. Never pick a model list_models marks "kind: jev" or "kind: laya" \u2014 those are text classification models and cannot run a routine.` },
         agentId: { type: "string", description: 'Required when flavor is "chat": an existing Customize -> Agents persona id.' },
         workspacePath: { type: "string", description: 'Required when flavor is "code": absolute path to the workspace directory.' },
         // Driven off schema.ts's list, not a second hand-maintained copy: this enum is what a MODEL