feat: add --provider-only flag to pin OpenRouter routing

Restricts OpenRouter routing to a specific provider (e.g. 'moonshotai'
for the official Kimi host) with fallbacks disabled, so a request fails
and is retried rather than silently routing to a third-party reseller
that may serve a different quantization.

Threaded through config["provider_only"] to both the initial call and
the re-prompt call. Adds two TDD tests for the request-body wiring.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -135,6 +135,12 @@ Override the sampling temperature from the config:
 uv run python src/benchmark_runner.py --model "openai/gpt-5.5" --temperature 0.7
 ```
 
+Pin OpenRouter routing to a specific provider — e.g. the model's official host instead of a third-party reseller. Fallbacks are disabled, so a request fails (and is retried) rather than silently routing elsewhere. Useful when resellers serve a different quantization than the official endpoint:
+
+```bash
+uv run python src/benchmark_runner.py --model "moonshotai/kimi-k2.6" --exam_name JEE_ADVANCED --exam_year 2026 --provider-only moonshotai
+```
+
 Resume an interrupted run (skips already-completed questions):
 
 ```bash

src/benchmark_runner.py

@@ -391,6 +391,7 @@ def process_question(
         request_timeout=config.get("request_timeout", 60),
         temperature=config.get("temperature", 0),
         reasoning_effort=config.get("reasoning_effort", "high"),
+        provider_only=config.get("provider_only"),
     )
 
     api_success = raw_response is not None  # False if API returned empty content
@@ -410,6 +411,7 @@ def process_question(
                 request_timeout=config.get("request_timeout", 60),
                 temperature=config.get("temperature", 0),
                 reasoning_effort=config.get("reasoning_effort", "high"),
+                provider_only=config.get("provider_only"),
             )
             if isinstance(parsed_answer_rp, list):
                 parsed_answer_rp = [str(item) for item in parsed_answer_rp]
@@ -1228,6 +1230,14 @@ if __name__ == "__main__":
         default=1,
         help="Number of independent runs per model (default: 1). Use 3+ for publication-grade variance analysis."
     )
+    parser.add_argument(
+        "--provider-only",
+        type=str,
+        default=None,
+        help="Optional: restrict OpenRouter routing to a specific provider (e.g. 'moonshotai' "
+             "for the official Kimi host). Comma-separated for multiple. Disables fallbacks, so "
+             "the call fails rather than routing to a reseller."
+    )
     args = parser.parse_args()
 
     # Dynamically update config path if user provides a different one
@@ -1263,6 +1273,10 @@ if __name__ == "__main__":
             if args.temperature is not None:
                 config["temperature"] = args.temperature
 
+            # Apply CLI provider-only override (restrict OpenRouter routing)
+            if args.provider_only:
+                config["provider_only"] = [p.strip() for p in args.provider_only.split(",") if p.strip()]
+
             if args.model not in config.get("openrouter_models", []):
                 logging.error(f"The model '{args.model}' is not listed in '{args.config}'. Please check the model name or the config file.")
                 exit(1)

src/llm_interface.py

@@ -104,6 +104,7 @@ def get_openrouter_prediction(
     request_timeout: int = 60,
     temperature: float = 0.0,
     reasoning_effort: str = "high",
+    provider_only: str | list[str] | None = None,
 ) -> tuple[list[str] | str | None, str | None, dict | None]:
     """
     Gets a prediction from an OpenRouter model. Handles initial image prompts and text-only re-prompts.
@@ -176,6 +177,12 @@ def get_openrouter_prediction(
             # - z-ai GLM-V: reasoning toggle
             data["reasoning"] = {"effort": reasoning_effort}
 
+        if provider_only:
+            # Restrict OpenRouter routing to specific provider(s) (e.g. the model's
+            # official host) and fail rather than silently fall back to a reseller.
+            only = [provider_only] if isinstance(provider_only, str) else list(provider_only)
+            data["provider"] = {"only": only, "allow_fallbacks": False}
+
         request_start_time = time.time()
         response = requests.post(
             OPENROUTER_API_ENDPOINT,

tests/test_llm_interface.py

@@ -90,3 +90,58 @@ def test_non_claude_model_no_thinking_param():
 
     assert "thinking" not in captured["json"]
     assert captured["json"]["temperature"] == 0.0
+
+
+# --- provider_only: restrict OpenRouter routing to a specific provider ---
+
+def test_provider_only_sets_provider_routing():
+    """When provider_only is given, the request restricts routing to that provider
+    with fallbacks disabled."""
+    captured = {}
+    def fake_post(url, headers, json, timeout):
+        captured["json"] = json
+        raise RuntimeError("stop after capture")
+
+    with patch("llm_interface.requests.post", side_effect=fake_post):
+        try:
+            from llm_interface import get_openrouter_prediction
+            get_openrouter_prediction(
+                model_identifier="moonshotai/kimi-k2.6",
+                api_key="test",
+                image=_make_test_image(),
+                exam_name="JEE_ADVANCED", exam_year="2026",
+                question_type="MCQ_SINGLE_CORRECT",
+                max_tokens=10000,
+                provider_only="moonshotai",
+            )
+        except RuntimeError:
+            pass
+
+    assert captured["json"]["provider"] == {
+        "only": ["moonshotai"],
+        "allow_fallbacks": False,
+    }
+
+
+def test_no_provider_only_omits_provider_routing():
+    """Without provider_only, no provider routing is sent (OpenRouter default)."""
+    captured = {}
+    def fake_post(url, headers, json, timeout):
+        captured["json"] = json
+        raise RuntimeError("stop after capture")
+
+    with patch("llm_interface.requests.post", side_effect=fake_post):
+        try:
+            from llm_interface import get_openrouter_prediction
+            get_openrouter_prediction(
+                model_identifier="moonshotai/kimi-k2.6",
+                api_key="test",
+                image=_make_test_image(),
+                exam_name="JEE_ADVANCED", exam_year="2026",
+                question_type="MCQ_SINGLE_CORRECT",
+                max_tokens=10000,
+            )
+        except RuntimeError:
+            pass
+
+    assert "provider" not in captured["json"]