From 00a9df8e9725518489050e02bc12ef83c9feeb80 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Fri, 12 Jun 2026 06:12:28 +0000
Subject: [PATCH] fix(translation): azure family detection uses v1's truthiness
 base_model-or-model (critic-azure M1, verifier-azure S1)

detection_model kept an empty-string base_model, so base_model: '' in a
deployment's litellm_params disabled the o-series/gpt-5 family and
response_format gates (fail-open: v2 served max_tokens where v1 routes
AzureOpenAIO1Config and emits max_completion_tokens). One-token fix to
'deps.base_model or model'; the docstring's 'exactly like v1' claim is
now true. Pinning corpus row uses base_model='' with the non-prefix
family substring deployment name prod-o3-mini per verifier-azure (the
shared openai prefix gate cannot catch it); row fails without the fix.
---
 litellm/translation/providers/azure/params.py   |  5 ++++-
 .../test_differential_azure_request.py          | 17 +++++++++++++++++
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/litellm/translation/providers/azure/params.py b/litellm/translation/providers/azure/params.py
index 6ed27c6ec9..8f76359ec0 100644
--- a/litellm/translation/providers/azure/params.py
+++ b/litellm/translation/providers/azure/params.py
@@ -30,7 +30,10 @@ _RESPONSE_FORMAT_SUPPORTED_MONTH = 8
 
 
 def detection_model(model: str, deps: TranslationDeps) -> str:
-    return deps.base_model if deps.base_model is not None else model
+    # v1 is truthiness, not None-ness (azure.py:246 `base_model or model`):
+    # base_model: "" in a config must fall to the deployment name, or the
+    # family/response_format gates fail OPEN (verifier-azure S1).
+    return deps.base_model or model
 
 
 def unsupported_model_family(model: str, deps: TranslationDeps) -> str | None:
diff --git a/tests/test_litellm/translation/test_differential_azure_request.py b/tests/test_litellm/translation/test_differential_azure_request.py
index f4b6f4419e..e85c099885 100644
--- a/tests/test_litellm/translation/test_differential_azure_request.py
+++ b/tests/test_litellm/translation/test_differential_azure_request.py
@@ -285,6 +285,23 @@ EXPECTED_FALLBACKS = {
         "o1-mini",
         "AzureOpenAIO1Config",
     ),
+    "o_series_empty_base_model_falls_to_deployment": (
+        # base_model: "" (a one-character YAML slip) must behave like v1's
+        # `base_model or model`: detection falls to the deployment name,
+        # whose NON-PREFIX o3 substring only azure's substring predicate
+        # catches (the shared openai prefix gate misses it). Proven divergent
+        # before the fix: v2 served max_tokens while v1 routed
+        # AzureOpenAIO1Config and emitted max_completion_tokens
+        # (verifier-azure S1).
+        {
+            "model": "prod-o3-mini",
+            "max_tokens": 100,
+            "messages": [{"role": "user", "content": "hi"}],
+        },
+        DEFAULT_API_VERSION,
+        "",
+        "AzureOpenAIO1Config",
+    ),
     "gpt5_model": (
         {"model": "gpt-5", "messages": [{"role": "user", "content": "hi"}]},
         DEFAULT_API_VERSION,