Merge pull request #6488 from happy5318/fix/qwen-prefixed-reasoning-heuristic

fix: Qwen reasoning heuristic fails for prefixed model IDs

api/config.py

@@ -3316,8 +3316,20 @@ def _candidate_supports_reasoning(candidate: str) -> bool:
             if major >= 4 or (major == 3 and minor >= 7):
                 return True
         return False
+    # Positive-only prefixed Qwen 3+ detection (e.g. "al-qwen3-8-max-preview"
+    # → tokens ["al","qwen3","8",...]). Scan for any token starting with
+    # "qwen" followed by version >= 3 and immediately allow. Do NOT return
+    # False here — Qwen 2.x embedded in hybrid IDs like
+    # "deepseek-r1-distill-qwen2.5-bakeneko-32b" must fall through to the
+    # DeepSeek detector below.
+    for token in tokens:
+        m = re.match(r"qwen(\d+)", token)
+        if m and int(m.group(1)) >= 3:
+            return True
+    # Terminal guard for standalone/bare Qwen IDs (original behavior):
+    # "qwen" as a standalone token or normalized starting with "qwen" means
+    # this IS a Qwen model — apply the 3+ gate and block 2.x.
     if "qwen" in token_set or normalized.startswith("qwen"):
-        # Restrict to Qwen 3+ (exclude Qwen 2/2.5)
         match = re.search(r"qwen.*?(\d+)(?:\D+(\d+))?", normalized)
         if match:
             major = int(match.group(1))

tests/test_reasoning_effort_model_capabilities.py

@@ -514,3 +514,50 @@ def test_datestamped_claude3_not_reasoning_capable_heuristic():
             f"{model} must remain reasoning-capable"
         )
 
+
+def test_qwen_prefixed_alias_reasoning_detection():
+    """Prefixed Qwen IDs (e.g. New-API aliases) must still be detected.
+
+    Regression: "al-qwen3.8-max-preview" normalizes to tokens
+    ["al", "qwen3", "8", ...] where "qwen" is NOT a standalone token,
+    so the old exact-membership check silently failed.
+    """
+    # Prefixed Qwen 3+ → reasoning-capable
+    for model in (
+        "al-qwen3.8-max-preview",
+        "al-qwen3.7-max",
+        "al-qwen3.7-plus",
+        "al-qwen3.6-flash",
+        "sn-qwen3-235b-a22b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is True, (
+            f"{model} must be reasoning-capable (prefixed Qwen 3+)"
+        )
+    # Bare Qwen 3+ → still works
+    for model in (
+        "qwen3-235b-a22b",
+        "qwen3-32b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is True, (
+            f"{model} must be reasoning-capable (bare Qwen 3+)"
+        )
+    # Qwen 2.x → excluded regardless of prefix
+    for model in (
+        "al-qwen2.5-72b-instruct",
+        "qwen2.5-7b-instruct",
+        "qwen2-72b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is False, (
+            f"{model} must NOT be reasoning-capable (Qwen 2.x excluded)"
+        )
+    # Hybrid IDs with embedded Qwen 2.x must NOT be shadowed by the Qwen
+    # branch — they must fall through to the DeepSeek detector.
+    for model in (
+        "deepseek-r1-distill-qwen2.5-bakeneko-32b",
+        "rinna/deepseek-r1-distill-qwen2.5-bakeneko-32b",
+    ):
+        assert cfg._candidate_supports_reasoning(model) is True, (
+            f"{model} must remain reasoning-capable (DeepSeek-R1 hybrid, "
+            f"Qwen 2.x must not shadow the DeepSeek detector)"
+        )
+