ur-agent 1.84.7 → 1.85.1

package/CHANGELOG.md

@@ -1,5 +1,54 @@
 # Changelog
 
+## 1.85.1
+
+- Added complete Anthropic workspace routing for identity-linked Claude API
+  keys. A validated workspace ID can be configured with
+  `anthropic.workspace_id`, `ANTHROPIC_WORKSPACE_ID`, or
+  `ur connect anthropic-api --workspace-id`; UR applies the matching header to
+  model discovery, diagnostics, streaming and non-streaming Messages requests,
+  and token counting while leaving workspace-scoped keys unchanged.
+- Preserved Anthropic's real missing-workspace response and replaced the
+  misleading provider/model compatibility failure with an actionable setup
+  instruction. Live model caches are partitioned by workspace so one workspace
+  cannot reuse another workspace's catalog.
+- Made OpenAI permanent account failures fail immediately instead of looking
+  like slow inference. Streaming HTTP error bodies are decoded safely, exact
+  provider error codes such as `billing_not_active` and `insufficient_quota`
+  are treated as non-retryable, and genuine transient rate limits continue to
+  use retry and backoff behavior.
+- Isolated provider-discovery tests from live keychain credentials and updated
+  public, configuration, troubleshooting, validation, provider, and technical
+  documentation for both behaviors.
+
+## 1.85.0
+
+- Split NVIDIA Build into two truthful provider surfaces backed by one saved
+  `NVIDIA_API_KEY`: **NVIDIA Agentic** for continuous tool-capable agent models
+  and **NVIDIA Special** for focused inference such as image, video, audio,
+  vision, retrieval, healthcare, climate, and optimization tasks. The generated
+  catalog currently mirrors all 36 Build cards marked Free Endpoint: 13
+  Agentic and 23 Special.
+- Rebuilt NVIDIA routing from each card's published inference contract. Every
+  executable model now uses its own documented HTTP endpoint and method or its
+  exact NVCF gRPC package, service, method, metadata, and streaming shape; direct
+  NVCF functions, Assets uploads, asynchronous polling, JSON-schema validation,
+  and binary artifact decoding are handled natively instead of being forced
+  through the shared chat-completions URL.
+- Added native execution for NVIDIA's five current public gRPC contracts,
+  including active-speaker detection, background-noise removal, Magpie TTS,
+  Studio Voice, and synthetic-video detection. Model descriptions expose their
+  purpose, expected input and output, transport, and contract before execution.
+- Made the official catalog non-destructive: account or invocation errors no
+  longer remove models. A card without a published API or proto contract remains
+  visible and explicitly unavailable rather than receiving an invented route.
+  Custom NVIDIA enterprise gateways retain their configurable endpoint and live
+  model discovery independently of the public Build catalog.
+- Updated provider diagnostics, `/model`, the `NvidiaSpecial` tool, public and
+  technical documentation, generated-catalog tooling, and regression coverage
+  for exact endpoint parity, shared credentials, non-pruning behavior, native
+  gRPC descriptors, assets, async results, artifacts, and focused-task dispatch.
+
 ## 1.84.7
 
 - Replaced NVIDIA's three-model one-shot allowlist with a generated catalog

package/README.md

@@ -289,6 +289,7 @@ ur config set provider anthropic-api
 ur config set provider gemini-api
 ur config set provider openrouter
 ur config set provider nvidia-nim
+ur config set provider nvidia-special
 ur config set provider unsloth
 ur config set model qwen2.5-coder:7b
 ur provider select-model ollama qwen2.5-coder:7b --json
@@ -298,6 +299,7 @@ ur config set provider.fallback ollama
 ur config set openrouter.routing auto
 ur config set openrouter.preferred_min_throughput 40
 ur config set openrouter.preferred_max_latency 3
+ur config set anthropic.workspace_id wrkspc_...
 ```
 
 `provider.fallback` records a recovery provider for `ur provider doctor`
@@ -308,7 +310,7 @@ select the recovery provider explicitly with `ur config set provider <id>`.
 provider before the URL to configure it without switching first, for example
 `ur config set base_url llama.cpp http://localhost:9931/v1`. UR remembers each
 provider's address independently, so switching among Ollama, LM Studio,
-llama.cpp, vLLM, Unsloth, NVIDIA NIM, or another compatible endpoint restores that
+llama.cpp, vLLM, Unsloth, NVIDIA Agentic, or another compatible endpoint restores that
 provider's last URL automatically. Existing single-URL settings are migrated
 to the previously active provider on the first provider switch or scoped
 base-URL write.
@@ -331,6 +333,11 @@ store records identifiers and status only. Compacted context is persisted only
 when `UR_OPENAI_RESPONSES_STATE_KEY` contains a 32-byte encryption key. Return
 to the default with `ur config set openai_transport chat-completions`.
 
+Both OpenAI transports retry genuine temporary rate limits. Permanent
+machine-coded account or billing responses such as `billing_not_active` and
+`insufficient_quota` are reported immediately, so an inactive project does not
+look like minute-long model latency.
+
 OpenRouter `auto` routing leaves tool turns to Auto Exacto and optimizes
 non-tool turns for end-to-end throughput. The routing, fallback,
 strict-parameter, rolling performance preference, service-tier, and supported
@@ -350,6 +357,21 @@ ur config set openrouter.speed fast            # standard | fast
 Direct Anthropic requests automatically preserve UR's prompt-cache
 breakpoints and enable per-tool fine-grained input streaming, reducing repeat
 prefill work and exposing large tool arguments as Claude generates them.
+Anthropic identity-linked keys that can act across workspaces also require a
+workspace on discovery, inference, streaming, and token-count calls. Configure
+the Console workspace ID once (it is not a secret):
+
+```sh
+ur config set anthropic.workspace_id wrkspc_...
+# or for one environment:
+export ANTHROPIC_WORKSPACE_ID=wrkspc_...
+```
+
+UR sends `anthropic-workspace-id` only when configured. Workspace-scoped keys
+continue to work without it. Use `auto` to clear the saved selection. If the
+key and workspace are being connected together, `ur connect anthropic-api
+--workspace-id wrkspc_...` accepts the same value.
+
 Anthropic's premium research-preview fast tier is opt-in and is sent only for
 Claude Opus 5 or Opus 4.8:
 
@@ -361,7 +383,8 @@ The account must have Anthropic fast-mode access. Unsupported Claude models
 stay on standard speed instead of receiving a fabricated provider option.
 
 Provider config accepts canonical IDs and common aliases. Examples:
-`openai-api`, `anthropic-api`, `gemini-api`, `openrouter`, `nvidia-nim` (`NVIDIA Build`), `ollama`,
+`openai-api`, `anthropic-api`, `gemini-api`, `openrouter`, `nvidia-nim` (`NVIDIA Agentic`),
+`nvidia-special` (`NVIDIA Special`), `ollama`,
 `lmstudio`, `LM Studio`, `llama.cpp`, `vllm`, `unsloth` (`Unsloth Studio`), and the subscription CLIs
 `codex-cli` (`chatgpt`), `claude-code-cli` (`claude`), `gemini-cli` (`gemini`),
 and `antigravity-cli` (`agy`). Use quotes for shell values with spaces.
@@ -378,10 +401,11 @@ ur connect logout openai-api          # clear a stored key
 | Provider | Access type | Runtime kind | Legal path |
 | --- | --- | --- | --- |
 | OpenAI API | API key | UR-native | `OPENAI_API_KEY` or `ur connect openai-api` |
-| Claude API | API key | UR-native | `ANTHROPIC_API_KEY` or `ur connect anthropic-api` |
+| Claude API | API key | UR-native | `ANTHROPIC_API_KEY` or `ur connect anthropic-api`; identity-linked keys also use `ANTHROPIC_WORKSPACE_ID` or `anthropic.workspace_id` |
 | Gemini API | API key | UR-native | `GEMINI_API_KEY` or `ur connect gemini-api` |
 | OpenRouter | API/router | UR-native | `OPENROUTER_API_KEY` or `ur connect openrouter` |
-| NVIDIA NIM | hosted/server API | UR-native | `NVIDIA_API_KEY` or `ur connect nvidia-nim`; configurable `base_url` |
+| NVIDIA Agentic | hosted/server API | UR-native | `NVIDIA_API_KEY` or `ur connect nvidia-nim`; configurable `base_url` |
+| NVIDIA Special | hosted/focused-task API | UR-native | the same `NVIDIA_API_KEY`; exact per-card HTTP/NVCF/gRPC inference contract |
 | Ollama | local/server | UR-native | configurable local, LAN, or hosted endpoint; optional `OLLAMA_API_KEY` |
 | LM Studio | local/server | UR-native | configurable endpoint; optional `LMSTUDIO_API_KEY` |
 | llama.cpp | local/server | UR-native | configurable endpoint; optional `LLAMA_CPP_API_KEY` |
@@ -404,30 +428,30 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    by source: `live` (discovered from the endpoint), `cache` (last discovery),
    `static` (predefined), or `unavailable` after a failed discovery with no
    fallback. Local/server providers (Ollama, LM Studio,
-   llama.cpp, vLLM, Unsloth) and OpenAI-compatible endpoints are discovered live. Hosted NVIDIA NIM
-   intersects its authenticated `/v1/models` response with UR's audited positive
-   agent contracts, so presence in NVIDIA's mixed inventory never makes an
-   embedding, parser, VLM, generator, or download-only card an ongoing chat model.
-   The separate NVCF deployment inventory does not narrow the hosted list. API
+   llama.cpp, vLLM, Unsloth) and OpenAI-compatible endpoints are discovered live. Public
+   NVIDIA Agentic discovery is generated from each current Build Free Endpoint
+   card that explicitly advertises agent/tool use; NVIDIA's account `/v1/models`
+   inventory does not narrow it, and a runtime entitlement error never removes a
+   model. A configured enterprise/self-hosted NIM gateway continues to use that
+   gateway's live `/models` response. API
    providers use live discovery from their `/models` endpoint once a key is
    connected (with a curated fallback list before that). Subscription CLIs show
    their curated model list because the official CLIs expose no models API. The
    generic `subscription` entry is an internal placeholder hidden from listings.
 
-   NVIDIA is split into two visibly labelled modes. `AGENT` models own the
-   ongoing tool-calling conversation. `ONE-SHOT` models run one specialized
-   job and never replace that agent. The latter come from a checked-in catalog
-   generated from NVIDIA's current public OpenAPI indexes, not from the chat-only
-   `/v1/models` feed. This release implements 92 exact task contracts spanning
-   text/image/video/3D generation, visual analysis, embeddings, reranking,
-   parsing, safety, translation, biology, molecular modeling, medical imaging,
-   route optimization, and climate simulation. Broken, staging-only,
-   status-only, download-only, and undocumented operations stay absent.
-   Focusing a task model shows what it is for; Enter remembers it for the next
-   matching NVIDIA task. UR validates its documented schema, routes to the exact
-   `integrate`, `ai`, `health`, `optimize`, or `climate` host, uploads large or
-   UUID-based files with NVIDIA Assets, polls asynchronous work, and writes
-   binary/large JSON results under `.ur/artifacts/nvidia/`.
+   NVIDIA is split into two provider entries. NVIDIA Agentic owns the ongoing
+   tool-calling conversation. NVIDIA Special selects one focused task and never
+   replaces that agent. The checked-in catalog is regenerated from all 100
+   currently visible Build cards and preserves all 36 cards labelled Free
+   Endpoint: 13 Agentic and 23 Special. Thirty-five publish executable
+   inference contracts; the one card whose invocation protocol is unpublished
+   remains visible and is labelled accordingly instead of being guessed or
+   removed. Every row shows its purpose, accepted input, and produced output.
+   Each executable Special model uses the URL, HTTP/RPC method, function ID,
+   request schema, and response schema in that model's own inference reference.
+   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.
 
    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
@@ -477,7 +501,7 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    provider's address untouched.
 
 Model lists never cross providers: OpenAI API, Claude API, Gemini API,
-OpenRouter, NVIDIA NIM, Ollama, and OpenAI-compatible local/server endpoints are separate
+OpenRouter, NVIDIA Agentic, NVIDIA Special, Ollama, and OpenAI-compatible local/server endpoints are separate
 access paths. API keys, local runtimes, and subscription logins are not
 interchangeable. The provider/model pair is validated before it is saved and
 again before every request; changing provider clears an incompatible model.
@@ -495,12 +519,11 @@ identity line in the system prompt reflects it too:
   `x-api-key` + `anthropic-version` on `/v1/messages`, OpenAI `Bearer` on
   `/v1/chat/completions` by default or `/v1/responses` when explicitly
   selected, Gemini `x-goog-api-key` on `:generateContent`, OpenRouter on its
-  OpenAI-compatible chat endpoint, and NVIDIA NIM agent models on their exact
+  OpenAI-compatible chat endpoint, and NVIDIA Agentic models on their exact
   documented hosted chat endpoint or a user-selected compatible NIM gateway.
-  NVIDIA one-shot models use their documented model endpoint across NVIDIA's
-  AI, retrieval, health, optimization, and climate APIs with the same stored
-  `NVIDIA_API_KEY`; asynchronous jobs are polled through NVIDIA's request ID
-  until completion or user cancellation.
+  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.
 - **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
@@ -517,7 +540,7 @@ identity line in the system prompt reflects it too:
   to choose a connected local, server, or API provider.
 
 Image-bearing tool results use each UR-native provider's valid multimodal wire
-shape. OpenAI Chat Completions, OpenRouter, NVIDIA NIM, LM Studio, llama.cpp,
+shape. OpenAI Chat Completions, OpenRouter, NVIDIA Agentic, LM Studio, llama.cpp,
 vLLM, Unsloth, and generic compatible endpoints keep the tool response textual and
 place its image in the immediately following user turn; Gemini nests the image
 parts in its function response; OpenAI Responses, Anthropic, and Ollama use their