ur-agent 1.81.6 → 1.81.8

package/CHANGELOG.md

@@ -1,5 +1,37 @@
 # Changelog
 
+## 1.81.8
+
+- Made reasoning effort provider- and model-accurate across the model picker,
+  `/effort`, the status bar, SDK, and outbound API requests. Up/Down now loads
+  the focused model's exact levels, Left/Right cycles only those levels, and
+  `max` resolves to the model's real `high`, `xhigh`, or `max` ceiling.
+- Added model-scoped llama.cpp reasoning discovery through `/props`, preserved
+  rich OpenAI-compatible capability metadata, and stopped presenting invented
+  graded levels when a provider does not advertise them.
+- Regenerated version-matched smoke and regression benchmark evidence (6/6 and
+  20/20 passed) so the release-readiness gate rejects no stale prior-release
+  artifacts.
+
+## 1.81.7
+
+- Made reasoning effort a real provider/model capability instead of a cosmetic
+  selector. UR now preserves exact `minimal`, `low`, `medium`, `high`, `xhigh`,
+  and `max` levels advertised by the provider and sends the selected value on
+  the provider request.
+- Made `/effort max` resolve to the selected model's actual ceiling. A
+  high-only model receives `high`, an xhigh model receives `xhigh`, and a model
+  with native max support receives `max`; `/effort status` and the status bar
+  report that same applied value.
+- Updated both model pickers so Up/Down refreshes the focused model's exact
+  effort list and Left/Right cycles only through supported levels. Models that
+  do not advertise graded reasoning no longer receive invented capabilities.
+- Added model-scoped llama.cpp reasoning discovery through `/props`, retained
+  rich OpenAI-compatible model metadata, and cached the result per endpoint and
+  model so clustered local deployments use their real capabilities.
+- Extended the CLI, SDK schemas, documentation, and regression coverage for
+  exact effort levels and provider wire serialization.
+
 ## 1.81.6
 
 - Restored the original clean welcome house without a shadow. The full adaptive

package/README.md

@@ -359,8 +359,13 @@ In the interactive app, `/model` is a two-step, provider-first picker:
    their curated model list because the official CLIs expose no models API. The
    generic `subscription` entry is an internal placeholder hidden from listings.
 
-   In the model catalog, use **Up/Down** to browse and **Left/Right** to change
-   the focused model's effort before pressing Enter. OpenRouter additionally
+   In the model catalog, use **Up/Down** to browse. The effort row updates to
+   the focused model's exact provider-advertised levels; use **Left/Right** to
+   cycle only those levels before pressing Enter. A generic `max` request is
+   resolved visibly to that model's actual `max`, `xhigh`, or `high` ceiling,
+   and the resolved value is the value sent to the provider. llama.cpp models
+   are checked lazily through their model-scoped `/props` capability while the
+   cursor moves. OpenRouter additionally
    shows compact model names, FREE/PAID tier, context size, tool/reasoning
    support, and the full untruncated ID immediately below the focused entry.
    Its catalog is fetched fresh every time it opens and never falls back to a