fix(agent): add nvidia/ to OpenRouter reasoning model prefixes (#75386)

[JonthanaHanh]

Summary

_supports_reasoning_extra_body() uses a hardcoded allowlist of OpenRouter model-family prefixes to decide whether to send reasoning.enabled: false in the API extra_body. nvidia/ was absent, so NVIDIA Nemotron reasoning models routed through OpenRouter never received the disable-reasoning signal — reasoning remained active (247-263 tokens/turn) even with explicit reasoning_effort: none.

Changes

• Added "nvidia/" to reasoning_model_prefixes in run_agent.py:6473
• Added regression test test_nvidia_models_are_treated_as_reasoning_capable covering three NVIDIA model variants

Test Plan

• pytest tests/run_agent/test_run_agent.py::TestSupportsReasoningExtraBody -xvs — 2 passed

Fixes #75386

[alt-glitch]

This was generated by AI during triage.

Duplicate of #75402: both patches add nvidia/ to the same OpenRouter reasoning prefix allowlist for #75386 and cover the same Nemotron behavior.

[teknium1]

Thanks for tracing the report to the OpenRouter gate. The current main path does still block the reporter's configuration: run_agent.py:6473-6484 lacks nvidia/, and agent/chat_completion_helpers.py:1337 passes that result to the OpenRouter profile.

Problems

• run_agent.py:6480 enables reasoning for every nvidia/ slug. The allowlist was introduced because reasoning configuration sent to NVIDIA via OpenRouter caused 400s (3f0f4a04a951bd5d64ba1f1c00be4f01945774fd). The issue establishes support for the reported Nemotron model, not the whole vendor namespace.
• tests/run_agent/test_run_agent.py:5688 checks only the boolean gate. It does not exercise the resulting request body, which is emitted by plugins/model-providers/openrouter/__init__.py:141-179.

Suggested changes

• Scope the new allowlist entry to model identifiers demonstrated to support the OpenRouter reasoning field, or establish a safe capability-based gate.
• Add a _build_api_kwargs() regression asserting the reporter's reasoning_effort: none becomes extra_body.reasoning = {"enabled": false}.

Automated hermes-sweeper review.

[teknium1]

This opts every nvidia/ model into extra_body.reasoning, but the original gate explicitly excluded NVIDIA after OpenRouter returned 400s for its reasoning payload (commit 3f0f4a04). Please restrict this to NVIDIA models known to accept the payload, or establish a safe capability gate rather than widening the entire vendor namespace.