ur-agent 1.85.1 → 1.85.2

package/CHANGELOG.md

@@ -1,5 +1,28 @@
 # Changelog
 
+## 1.85.2
+
+- Made NVIDIA Special selections directly executable from the next prompt. UR
+  now sends focused image, video, audio, vision, retrieval, and other one-shot
+  jobs through the selected model's exact NVIDIA inference contract without
+  invoking or depending on the ongoing Anthropic, OpenAI, Ollama, or other
+  conversation provider. Structured prompt, file, passage, output-path, and
+  exact-JSON inputs are accepted, and missing required media produces targeted
+  guidance instead of an unrelated provider error.
+- Restored reliable model switching across providers. Provider and model now
+  change atomically, selecting an ordinary model exits NVIDIA Special mode,
+  provider-scoped models take precedence over stale legacy global settings,
+  and a saved model from one backend can no longer be paired accidentally with
+  another backend.
+- Added NVIDIA Special to first-run and new-folder provider selection while
+  keeping a normal agent model available for the conversation, and made the
+  active focused-task model visible in the status line. Updated public,
+  troubleshooting, provider, validation, and technical documentation, with
+  end-to-end regression coverage proving focused tasks bypass the conversation
+  provider and stale provider/model pairs recover correctly.
+- Refreshed the locked `fast-uri` dependency to its compatible 3.1.6 release so
+  the production dependency audit remains clean without changing runtime policy.
+
 ## 1.85.1
 
 - Added complete Anthropic workspace routing for identity-linked Claude API

package/README.md

@@ -452,6 +452,13 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    UR supports direct HTTP, model-specific NVCF invocation, NVIDIA Assets,
    asynchronous polling, and the five documented Maxine/Riva gRPC services;
    artifacts are written under `.ur/artifacts/nvidia/` unless a path is supplied.
+   Selecting a Special row activates a real task mode: the next non-command
+   prompt goes directly to that card's inference endpoint and never passes
+   through the ongoing Anthropic, OpenAI, Ollama, or other chat provider. Plain
+   text becomes `prompt`. Models requiring media accept newline fields such as
+   `video_path: /path/source.mp4` or `image_path: /path/image.png`; exact JSON is
+   also accepted. First-workspace setup shows NVIDIA Special too, then returns
+   to the provider list so an ordinary agent model can be chosen for other work.
 
    In the model catalog, use **Up/Down** to browse. For graded models, the effort row updates to
    the focused model's capability-backed selectors; use **Left/Right** to cycle
@@ -523,7 +530,9 @@ identity line in the system prompt reflects it too:
   documented hosted chat endpoint or a user-selected compatible NIM gateway.
   NVIDIA Special uses each card's documented HTTP, direct NVCF, or gRPC
   endpoint with the same stored `NVIDIA_API_KEY`; asynchronous jobs are polled
-  through NVIDIA's request ID until completion or user cancellation.
+  through NVIDIA's request ID until completion or user cancellation. While a
+  Special task is selected, the next plain prompt is dispatched directly and
+  does not require the ongoing agent provider to be connected.
 - **Local/server** providers call the configured endpoint (`/v1/chat/completions`
   for LM Studio/llama.cpp/vLLM/Unsloth; the native API for Ollama). Unsloth is
   provider-only: UR never starts, installs, updates, trains, or loads models in