ur-agent 1.84.3 → 1.84.5

package/CHANGELOG.md

@@ -1,5 +1,76 @@
 # Changelog
 
+## 1.84.5
+
+- Fixed a provider-wide multi-turn tool-history failure. Some
+  OpenAI-compatible models reuse response-scoped IDs such as `TaskCreate:0`;
+  UR previously classified the later completed pair as a duplicate, removed
+  it from the API-bound history, and could prompt the model to say the user's
+  answer was empty. Later unambiguous call/result pairs now receive stable,
+  conversation-unique IDs before pairing repair, while the saved transcript
+  remains unchanged and genuinely corrupt duplicates still fail normally.
+- Restored NVIDIA Build's full hosted chat catalog by making its authenticated
+  `/v1/models` response authoritative. UR no longer intersects hosted models
+  with the unrelated NVCF deployment-function inventory that hid valid
+  endpoints; non-agent utility endpoints remain filtered, and download-only
+  Build cards are never inserted into the hosted picker. Doctor and runtime
+  invalidation now follow the same catalog contract.
+- Reduced NVIDIA request overhead and added exact model-native behavior. Hosted
+  NIM token analysis now uses the local request-shaped estimate immediately
+  instead of probing an unsupported count route. Nemotron 3.5 Lightning is
+  preferred when NVIDIA returns it and its native on/off thinking field is
+  wired; Kimi K3 retains its mandatory documented effort contract, and no
+  unknown model inherits a fabricated level or Ultra.
+- Added documented direct-Anthropic latency controls. Prompt-cache breakpoints
+  now survive native request translation, streaming tools use per-tool
+  fine-grained input streaming, and `anthropic.speed=fast` opts enabled Opus
+  5/4.8 accounts into Anthropic's premium research-preview tier with the
+  required beta header. Unsupported models stay on standard speed and response
+  accounting preserves the tier actually served.
+- Audited other provider acceleration paths against their current APIs.
+  OpenAI Responses already retains native SSE/WebSocket continuation, Gemini
+  implicit caching remains automatic, and UR does not invent Interactions-only
+  Gemini Priority fields or OpenRouter-style routing controls for NVIDIA and
+  local servers. User, troubleshooting, validation, provider, and technical
+  documentation now match these executable boundaries.
+
+## 1.84.4
+
+- Corrected NVIDIA Build model discovery so the picker no longer treats the
+  hosted `/v1/models` feed as proof that a model can be invoked. UR now
+  intersects that feed with the connected account's authenticated ACTIVE NVCF
+  functions and removes embedding, safety, parser, translation, retrieval,
+  reward, detector, and other non-agent endpoints. Custom enterprise and
+  self-hosted NIM gateways remain independent and use their own configured
+  catalogs.
+- Made NVIDIA diagnostics and runtime catalog churn actionable. `ur provider
+  doctor nvidia-nim` now validates the selected model against the account-active
+  inventory; a definitive missing-function 404 removes that model from the
+  endpoint-scoped session catalog and reports a model-selection fix without
+  displaying or retaining NVIDIA's internal account and function identifiers.
+- Added production regressions for retired hosted models, account inventory
+  failures, utility-model filtering, custom gateway isolation, selected-model
+  doctor failures, request/stream error redaction, and runtime cache
+  invalidation. User, troubleshooting, validation, configuration, and technical
+  documentation now describe the executable behavior.
+- Corrected the output-limit diagnostic: a `max_tokens`/`length` finish now
+  says that the provider stopped the selected model at a per-response boundary,
+  rather than incorrectly claiming that UR's accumulated response exceeded a
+  total limit. Long work now continues without a fixed continuation count while
+  each response makes novel progress; two consecutive empty/replayed capped
+  responses stop only the stalled loop. Practical response chunks preserve
+  router capacity and local KV memory, while the selected model's advertised
+  limit remains the upper bound for explicit overrides.
+- Updated OpenRouter dispatch to its current routing contract. Tool turns no
+  longer force `sort: latency`, which had disabled OpenRouter Auto Exacto;
+  they now use the router's live throughput, tool-call reliability, and
+  benchmark ordering. Non-tool turns default to end-to-end throughput rather
+  than time-to-first-token alone. Stable session affinity remains enabled for
+  prompt caching, and CLI settings now expose routing, fallback, parameter,
+  latency/throughput preference, service-tier, and supported fast-mode controls.
+  OpenRouter's `:nitro`, `:floor`, and `:exacto` virtual model variants now
+  validate against the discovered base model and inherit its capabilities.
+
 ## 1.84.3
 
 - Updated both production and release artifact uploads to the current official

package/README.md

@@ -295,6 +295,9 @@ ur provider select-model ollama qwen2.5-coder:7b --json
 ur config set base_url ollama http://localhost:11434
 ur config set base_url llama.cpp http://localhost:9931/v1
 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
 ```
 
 `provider.fallback` records a recovery provider for `ur provider doctor`
@@ -328,6 +331,35 @@ 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`.
 
+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
+fast-mode controls are configurable; `priority` and `fast` are optional and may
+cost more. Native `:nitro`, `:floor`, and `:exacto` model suffixes are accepted
+and keep the base model's discovered context, output, tool, and reasoning
+capabilities.
+
+```sh
+ur config set openrouter.routing auto          # auto | throughput | latency | price
+ur config set openrouter.allow_fallbacks true    # true | false | auto
+ur config set openrouter.require_parameters true # true | false | auto
+ur config set openrouter.service_tier priority # auto | default | flex | priority | fast
+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's premium research-preview fast tier is opt-in and is sent only for
+Claude Opus 5 or Opus 4.8:
+
+```sh
+ur config set anthropic.speed fast             # standard | fast
+```
+
+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`,
 `lmstudio`, `LM Studio`, `llama.cpp`, `vllm`, `unsloth` (`Unsloth Studio`), and the subscription CLIs
@@ -372,7 +404,9 @@ 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, including NVIDIA NIM, are discovered live; API
+   llama.cpp, vLLM, Unsloth) and OpenAI-compatible endpoints are discovered live. Hosted NVIDIA NIM
+   uses its documented `/v1/models` response directly and removes non-agent utility
+   endpoints before presenting a model; the separate NVCF deployment inventory does not narrow it. 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
@@ -382,8 +416,8 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    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
    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`
+   two-state mapping (currently Ollama, direct Anthropic, and NVIDIA Nemotron
+   3.5 Lightning), 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.
    `ultra` is UR's visible
@@ -412,8 +446,12 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    support, and the full untruncated ID immediately below the focused entry.
    Its endpoint-scoped catalog is reused for five minutes; Ctrl+R forces an
    immediate live refresh without substituting stale entries. Interactive
-   requests prefer OpenRouter's latency routing and reuse a stable session ID
-   and provider-authored prompt-cache markers for warmer multi-turn streams.
+   tool turns preserve OpenRouter Auto Exacto so the router can combine live
+   throughput with measured tool-call reliability; ordinary text turns prefer
+   end-to-end throughput rather than time-to-first-token alone. UR reuses a
+   stable session ID and provider-authored prompt-cache markers for warmer
+   multi-turn streams. Configure the policy with `openrouter.routing`, or use
+   the native `:nitro`, `:floor`, and `:exacto` model variants.
    API-key entry is masked, aligned on one row, and stored through the OS
    keychain flow.