Sync from GitHub (tests passed)

app/commentary.py

@@ -383,6 +383,7 @@ Write as a seasoned commodities strategist would for institutional clients—pre
         "fallback_models": settings.openrouter_fallback_models_list,
         "chain_deadline_seconds": getattr(settings, "openrouter_chain_deadline_seconds", 120.0),
         "provider": {"require_parameters": True},
+        "extra_payload": {"reasoning": {"effort": "medium"}},
         "referer": "https://copper-mind.vercel.app",
         "title": "CopperMind Commentary",
     }

app/openrouter_client.py

@@ -120,8 +120,14 @@ def _extract_message_text(data: dict[str, Any]) -> str:
 
 def _candidate_payload(base_payload: dict[str, Any], candidate: str) -> dict[str, Any]:
     payload = {**base_payload, "model": candidate}
-    if candidate.lower().startswith("nvidia/nemotron") and "reasoning" not in payload:
-        payload["reasoning"] = {"effort": "none"}
+    existing_reasoning = payload.get("reasoning")
+    if candidate.lower().startswith("nvidia/nemotron"):
+        if isinstance(existing_reasoning, dict) and "effort" in existing_reasoning:
+            payload["reasoning"] = {"effort": existing_reasoning["effort"]}
+        else:
+            payload["reasoning"] = {"effort": "low"}
+    elif isinstance(existing_reasoning, dict) and set(existing_reasoning.keys()) == {"effort"}:
+        payload.pop("reasoning", None)
     return payload
 
 
@@ -178,12 +184,19 @@ async def create_chat_completion(
                 if time.monotonic() >= deadline:
                     raise OpenRouterError("OpenRouter model chain deadline exceeded", category="timeout", model=candidate, attempts=attempts)
                 await _throttle_request(rpm)
+                remaining_before_call = deadline - time.monotonic()
+                if remaining_before_call <= 0:
+                    raise OpenRouterError("OpenRouter model chain deadline exceeded", category="timeout", model=candidate, attempts=attempts)
+                attempt_timeout = min(float(timeout_seconds), max(remaining_before_call, 0.1))
                 payload = _candidate_payload(base_payload, candidate)
                 response: Optional[httpx.Response] = None
                 try:
-                    response = await client.post("https://openrouter.ai/api/v1/chat/completions", headers=headers, json=payload)
-                except httpx.TimeoutException as exc:
-                    last_error = OpenRouterError(str(exc), category="timeout", model=candidate)
+                    response = await asyncio.wait_for(
+                        client.post("https://openrouter.ai/api/v1/chat/completions", headers=headers, json=payload),
+                        timeout=attempt_timeout,
+                    )
+                except (httpx.TimeoutException, asyncio.TimeoutError, TimeoutError) as exc:
+                    last_error = OpenRouterError(str(exc) or "OpenRouter request timed out", category="timeout", model=candidate)
                 except httpx.RequestError as exc:
                     last_error = OpenRouterError(str(exc), category="network", model=candidate)
                 else:
@@ -210,7 +223,14 @@ async def create_chat_completion(
 
                 status = response.status_code if response is not None else None
                 retryable = last_error is not None and last_error.category in retryable_categories
-                will_retry = retryable and attempt_index < max(0, int(max_retries))
+                # When fallback models exist, advance immediately on a timeout instead of
+                # burning a second 45s window on the same unresponsive upstream model.
+                skip_same_model_timeout_retry = (
+                    last_error is not None
+                    and last_error.category == "timeout"
+                    and len(models) > 1
+                )
+                will_retry = retryable and not skip_same_model_timeout_retry and attempt_index < max(0, int(max_retries))
                 attempts.append(_attempt(candidate, attempt_index + 1, last_error.category if last_error else "network", status, will_retry))
                 if last_error and last_error.category == "auth":
                     last_error.attempts = attempts

app/settings.py

@@ -92,15 +92,15 @@ class Settings(BaseSettings):
     # Deprecated - kept for backward compatibility. Keep this on a currently
     # listed free model so installations without an explicit env override do
     # not silently start with a retired OpenRouter slug.
-    openrouter_model: str = "nvidia/nemotron-3.5-super-vl:free"
+    openrouter_model: str = "nvidia/nemotron-3-super-120b-a12b:free"
     # Keep fast and reliable roles independent so one provider/model removal
     # cannot collapse the entire LLM path. NVIDIA and Liquid passed the exact
     # credentialed scoring/commentary probe on 2026-09-27; the Gemma free
     # models were temporarily rate-limited by their shared upstream pool.
-    openrouter_model_scoring: str = "nvidia/nemotron-3.5-super-vl:free"
-    openrouter_model_scoring_fast: Optional[str] = "nvidia/nemotron-3.5-super-vl:free"
+    openrouter_model_scoring: str = "nvidia/nemotron-3-super-120b-a12b:free"
+    openrouter_model_scoring_fast: Optional[str] = "nvidia/nemotron-3-super-120b-a12b:free"
     openrouter_model_scoring_reliable: Optional[str] = "liquid/lfm-2.5-2.6b:free"
-    openrouter_model_commentary: str = "nvidia/nemotron-3.5-super-vl:free"
+    openrouter_model_commentary: str = "nvidia/nemotron-3-super-120b-a12b:free"
     # Stay below the shared free-provider request ceiling. The worker still
     # uses bounded retries and a per-chain deadline for transient 429s.
     openrouter_rpm: int = 12
@@ -115,7 +115,8 @@ class Settings(BaseSettings):
     # client as transport-level fallbacks when the primary model 429s/5xx's.
     # Example: "google/gemini-flash-1.5:free,meta-llama/llama-3.1-8b-instruct:free"
     openrouter_fallback_models: Optional[str] = (
-        "liquid/lfm-2.5-2.6b:free,nvidia/nemotron-3.5-super-vl:free,"
+        "liquid/lfm-2.5-2.6b:free,nvidia/nemotron-3-super-120b-a12b:free,"
+        "nvidia/nemotron-3.5-super-vl:free,"
         "google/gemma-4-31b-it:free,google/gemma-4-26b-a4b-it:free"
     )
     tokenizers_parallelism: str = "false"
@@ -134,7 +135,7 @@ class Settings(BaseSettings):
     
     # LLM Sentiment Analysis
     # Deprecated - kept for backward compatibility
-    llm_sentiment_model: str = "nvidia/nemotron-3.5-super-vl:free"
+    llm_sentiment_model: str = "nvidia/nemotron-3-super-120b-a12b:free"
     
     # Pipeline trigger authentication
     pipeline_trigger_secret: Optional[str] = None
@@ -255,7 +256,7 @@ class Settings(BaseSettings):
                 self.llm_sentiment_model,
                 self.openrouter_model,
             )
-            or "nvidia/nemotron-3.5-super-vl:free"
+            or "nvidia/nemotron-3-super-120b-a12b:free"
         )
 
     @property
@@ -273,7 +274,7 @@ class Settings(BaseSettings):
                 self.llm_sentiment_model,
                 self.openrouter_model_scoring,
             )
-            or "nvidia/nemotron-3.5-super-vl:free"
+            or "nvidia/nemotron-3-super-120b-a12b:free"
         )
 
     @property

scripts/probe_openrouter_contracts.py

@@ -106,9 +106,9 @@ async def run_probe(*, fast: str, reliable: str, commentary_model: str, batch_si
 
 def main() -> int:
     parser = argparse.ArgumentParser(description="Probe exact OpenRouter production contracts")
-    parser.add_argument("--fast", default="nvidia/nemotron-3.5-super-vl:free")
+    parser.add_argument("--fast", default="nvidia/nemotron-3-super-120b-a12b:free")
     parser.add_argument("--reliable", default="liquid/lfm-2.5-2.6b:free")
-    parser.add_argument("--commentary", default="nvidia/nemotron-3.5-super-vl:free")
+    parser.add_argument("--commentary", default="nvidia/nemotron-3-super-120b-a12b:free")
     parser.add_argument("--batch-size", type=int, default=12, help="Synthetic scoring items, up to the worker's 12-item chunk")
     args = parser.parse_args()
     result = asyncio.run(run_probe(fast=args.fast, reliable=args.reliable, commentary_model=args.commentary, batch_size=args.batch_size))