ur-agent 1.84.0 → 1.84.2

package/CHANGELOG.md

@@ -1,5 +1,51 @@
 # Changelog
 
+## 1.84.2
+
+- Added NVIDIA NIM as a real UR-native provider for build.nvidia.com. It uses
+  NVIDIA's hosted `https://integrate.api.nvidia.com/v1` default, live model
+  discovery, Bearer authentication from `NVIDIA_API_KEY` or `ur connect
+  nvidia-nim`, configurable provider-scoped endpoints, streaming, tool calls,
+  multimodal input, NIM token counting with a local-estimate fallback, and an
+  optional live smoke target. Only NVIDIA-documented per-model effort ladders
+  are overlaid; future/unknown models do not inherit Ultra.
+- Made authentication fully usable on the generic `openai-compatible`
+  provider. `/connect openai-compatible` stores its optional key, and the
+  `/model` screen now exposes `K API key` to add or replace one without making
+  anonymous local endpoints key-required. Stored endpoints and credentials
+  remain provider-scoped across switches.
+- Added a real GNU-compatible `timeout` command inside UR's spawned macOS
+  shell when neither native `timeout` nor Homebrew `gtimeout` is available.
+  It uses the Bash tool's current Node/Bun runtime, supervises an isolated
+  process group, supports the common GNU duration/signal/kill-after flags, and
+  returns status 124 on expiry. The Bash prompt now prefers its native
+  millisecond timeout field, while generated Linux-style commands remain
+  executable on macOS instead of failing with `command not found: timeout`.
+- Made image-bearing tool results portable across every UR-native provider.
+  OpenAI Chat Completions, OpenRouter, NVIDIA NIM, LM Studio, llama.cpp, vLLM, Unsloth,
+  and generic compatible endpoints keep a textual `role: tool` result and send
+  the image bytes in the immediately following multimodal user turn; Gemini
+  nests the image parts in its function response; OpenAI Responses, Anthropic,
+  and Ollama retain their native rich-result paths. Added a provider matrix and
+  real shell deadline regressions so screenshots are neither rejected nor
+  silently discarded.
+
+## 1.84.1
+
+- Made dynamic reasoning discovery capability-truthful across local and
+  OpenAI-compatible providers. vLLM now uses its non-generating server-info
+  contract to expose `minimal→none|low|medium|high` only when a reasoning
+  parser is configured, and the adapter serializes all provider-authored effort
+  aliases rather than translating Ultra alone. llama.cpp records its template
+  support flag without fabricating a finite level set that the endpoint does
+  not publish.
+- Refined Ollama's thinking presentation: a generic `thinking` capability now
+  reports thinking support and UR's verified native on/off control without
+  claiming the model supports only booleans. Exact graded levels and Ultra
+  remain model-scoped and provider-advertised. Added provider-matrix, discovery,
+  picker, command-message, and request-wire regression coverage and synchronized
+  the public site, user guides, validation guide, and technical specification.
+
 ## 1.84.0
 
 - Added production npm marketplace sources with scoped-package and version

package/README.md

@@ -288,6 +288,7 @@ ur config set provider openai-api
 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 unsloth
 ur config set model qwen2.5-coder:7b
 ur provider select-model ollama qwen2.5-coder:7b --json
@@ -304,7 +305,7 @@ 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, or another compatible endpoint restores that
+llama.cpp, vLLM, Unsloth, NVIDIA NIM, 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.
@@ -328,7 +329,7 @@ when `UR_OPENAI_RESPONSES_STATE_KEY` contains a 32-byte encryption key. Return
 to the default with `ur config set openai_transport chat-completions`.
 
 Provider config accepts canonical IDs and common aliases. Examples:
-`openai-api`, `anthropic-api`, `gemini-api`, `openrouter`, `ollama`,
+`openai-api`, `anthropic-api`, `gemini-api`, `openrouter`, `nvidia-nim` (`NVIDIA Build`), `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.
@@ -348,6 +349,7 @@ ur connect logout openai-api          # clear a stored key
 | Claude API | API key | UR-native | `ANTHROPIC_API_KEY` or `ur connect anthropic-api` |
 | 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` |
 | 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` |
@@ -370,7 +372,7 @@ 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; API
+   llama.cpp, vLLM, Unsloth) and OpenAI-compatible endpoints, including NVIDIA NIM, are discovered live; 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
@@ -379,7 +381,8 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    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
    only values UR can map to provider-native levels before pressing Enter. For
-   boolean-thinking models on runtimes with a native two-state mapping (currently
+   models with thinking but no advertised graded ladder on runtimes with a native
+   two-state mapping (currently
    Ollama and direct Anthropic), Left selects off, Right selects on, and `t`
    toggles. The same state is available directly through `/thinking on|off`.
    Generic OpenAI-compatible runtimes never receive an invented boolean field.
@@ -387,14 +390,21 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    beyond-high ceiling selector. It appears only when the provider/model
    advertises `ultra`, `max`, `xhigh`, or an explicit equivalent, and the row
    shows the exact mapping (for example, `ultra→max`). Models whose graded
-   ladder tops out at `high`, plus boolean-thinking models, do not get Ultra. A generic
+   ladder tops out at `high`, plus models without an advertised beyond-high
+   value, do not get Ultra. A generic
    `max` request resolves visibly to that model's highest supported non-Ultra tier
    (commonly `max`, `xhigh`, or `high`), and that resolved value is sent to the provider. llama.cpp models
    are checked lazily through their model-scoped `/props` capability while the
-   cursor moves. Ollama models are checked through `/api/show`; its `thinking`
-   capability enables boolean thinking without inventing a graded ladder.
-   If `/effort max` or another graded request is used for a boolean-only model,
-   UR enables thinking but reports that no graded level was sent.
+   cursor moves. Because current llama.cpp reports support but not the accepted
+   level names, that flag alone does not create a graded selector. vLLM is
+   checked lazily through `/server_info?config_format=json`; a configured
+   reasoning parser enables its documented `none|low|medium|high` Chat
+   Completions contract (`minimal→none`) without inventing Ultra. Ollama models
+   are checked through `/api/show`; a generic `thinking` capability establishes
+   thinking support but not a model-specific ladder, so UR uses Ollama's native
+   on/off control unless the endpoint supplies exact levels.
+   If `/effort max` or another graded request is used for such a model, UR
+   enables thinking but reports that no graded level was sent.
    GPT-OSS uses its documented `low|medium|high` values, while any other graded
    values or Ultra aliases must be explicitly present in provider metadata.
    The resolved value is sent through Ollama's native `think` field. OpenRouter additionally
@@ -412,7 +422,7 @@ 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, Ollama, and OpenAI-compatible local/server endpoints are separate
+OpenRouter, NVIDIA NIM, 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.
@@ -429,8 +439,9 @@ identity line in the system prompt reflects it too:
 - **API** providers call each service in its native wire format — Anthropic
   `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`, and OpenRouter on its
-  OpenAI-compatible chat endpoint.
+  selected, Gemini `x-goog-api-key` on `:generateContent`, OpenRouter on its
+  OpenAI-compatible chat endpoint, and NVIDIA NIM on its official hosted or
+  user-selected OpenAI-compatible endpoint.
 - **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
@@ -446,6 +457,15 @@ identity line in the system prompt reflects it too:
   subscription backend is configured, `/model` marks it unavailable and asks you
   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,
+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
+native rich-result forms. The selected model must still support vision.
+External subscription CLIs remain text-only because their official prompt
+interfaces do not expose a portable binary-image channel.
+
 Ollama is used only when Ollama is selected. There is no silent cross-provider
 fallback: if dispatch fails, UR reports the selected provider, model, and runtime
 backend. Use `ur provider status` (or `ur provider doctor <id>`) to inspect the